Clinical Reasoning Under a Partially Observed Objective in Cone Beam CT Report Generation

# Clinical Reasoning Under a Partially Observed Objective in Cone Beam CT Report Generation
Source: [https://arxiv.org/html/2609.13238](https://arxiv.org/html/2609.13238)
Govind Arun[https://orcid.org/0009-0002-2573-4003](https://orcid.org/0009-0002-2573-4003)Affiliation:University of Maryland, College Park, Maryland, USASidharth N Krishna[https://orcid.org/0009-0002-4336-2477](https://orcid.org/0009-0002-4336-2477)Affiliation:Indira Gandhi National Open University, New Delhi, IndiaUma Ranjan[https://orcid.org/0000-0001-6258-4513](https://orcid.org/0000-0001-6258-4513)Affiliation:Indian Institute of Technology Jammu, India

###### Abstract

Maxillofacial report generation from cone beam computed tomography is scored here by a composite objective placing 80% of its weight on a large language model judgement of factual entailment and 20% on lexical overlap, of which only the lexical fifth is visible during development\. The grader’s BLEU\-4 and METEOR routines are reproduced in pure Python and match the reference to machine precision, and an offline entailment surrogate, which tells a report written for one patient from one written for another at an area under the curve of 0\.987, makes the composite objective cheap enough to optimise directly\. Over the 622 case public release, a report selected against the visible lexical ranking scores 0\.2909, whereas one selected against the composite objective scores 0\.4122, because pursuing n\-gram overlap drives entailment precision from 0\.522 down to 0\.266\. A 29 million parameter encoder fine tuned on the release reaches a prevalence weighted out of fold area under the curve of 0\.486 over 985 statements, indistinguishable from the corpus prior, while nine numbers read from the image header reach 0\.945 for mandible coverage and 0\.872 for condyle coverage, and acquisition centre alone predicts sentence choice at 0\.718 against 0\.663 for the image derived model, identifying dictation convention rather than anatomy as the quantity the lexical metrics reward\. The delivered system emits eight unconditional statements and five gated on header geometry under polarity, laterality and tooth level consistency constraints, and reaches METEOR 0\.3542 over 50 held out cases from an unseen centre\. The dataset and code are available at[https://github\.com/GIND123/CBCT\-Clinical\-Reasoner](https://github.com/GIND123/CBCT-Clinical-Reasoner)

###### Keywords:

Report generation Cone beam computed tomography Constrained decoding Factual entailment Objective misspecification

## 1Introduction

Three dimensional imaging is routine in dentistry and maxillofacial surgery, yet the report that turns a volume into a surgical decision is still dictated by hand\. Automating that step is now posed as a benchmark task: from one cone beam computed tomography \(CBCT\) volume of the jaws, produce a report covering dental status, bone quality, anatomical variants, proximity to critical structures and procedure related risk\[[6](https://arxiv.org/html/2609.13238#bib.bib1)\]\. The release extends a segmentation oriented series\[[8](https://arxiv.org/html/2609.13238#bib.bib2),[7](https://arxiv.org/html/2609.13238#bib.bib3)\]\[[8](https://arxiv.org/html/2609.13238#bib.bib2),[7](https://arxiv.org/html/2609.13238#bib.bib3),[11](https://arxiv.org/html/2609.13238#bib.bib18)\]with paired clinical text, moving the task from delineating anatomy to asserting facts about it\[[20](https://arxiv.org/html/2609.13238#bib.bib14)\]\.

Report generation is usually measured through lexical similarity, and the limits of that practice are well documented\[[18](https://arxiv.org/html/2609.13238#bib.bib6),[22](https://arxiv.org/html/2609.13238#bib.bib15)\]\. The benchmark studied here responds by weighting a large language model judgement of factual entailment, RadFact\[[4](https://arxiv.org/html/2609.13238#bib.bib4)\], at four times the combined weight of BLEU\-4 and METEOR, then withholds that component during development: the entailment judge is disabled on the evaluation platform, so the public ranking reflects only the lexical fifth of the score that decides the outcome\. Participants tune against a projection of the objective rather than the objective itself, and that projection turns out to be actively misleading\.

The proposed system is built around that observation, with the measurements that forced each decision\. The grader is reimplemented exactly, so the full objective becomes computable offline and selection runs against it rather than its visible projection, worth 0\.121 composite score \(Section[6](https://arxiv.org/html/2609.13238#S6)\)\. The recoverable signal is localised by measurement rather than assumed \(Section[4](https://arxiv.org/html/2609.13238#S4)\), decoding constraints are read off the arithmetic of the entailment metric so that self contradiction and paraphrase, both of which pay under lexical scoring, are structurally excluded \(Section[5](https://arxiv.org/html/2609.13238#S5)\), and the container is built so that no single load failure can silence the model, after an earlier submission lost roughly fourteen ranking positions to exactly that \(Section[7](https://arxiv.org/html/2609.13238#S7)\)\.

![Refer to caption](https://arxiv.org/html/2609.13238v1/system_design.png)Figure 1:Overview of the proposed CBCT report generation pipeline\. The system extracts image and acquisition signals, selects candidate clinical statements, enforces consistency constraints, and applies geometry\-based gating before assembling the final structured clinical report\.
## 2The Scored Objective

Entries are ranked by

Final=0\.8⋅F1RadFact\+0\.2⋅12​\(BLEU​\-​4\+METEOR\),\\mathrm\{Final\}\\;=\\;0\.8\\cdot F\_\{1\}^\{\\mathrm\{RadFact\}\}\\;\+\\;0\.2\\cdot\\tfrac\{1\}\{2\}\\left\(\\mathrm\{BLEU\\text\{\-\}4\}\+\\mathrm\{METEOR\}\\right\),\(1\)with the entailment term computed offline once the submission window closes\[[6](https://arxiv.org/html/2609.13238#bib.bib1)\]\. Three details in the released evaluation code shape everything downstream\. BLEU\-4 is aggregated at corpus level through NLTK\[[16](https://arxiv.org/html/2609.13238#bib.bib5),[5](https://arxiv.org/html/2609.13238#bib.bib12)\]with method one smoothing, so per case averaging yields a different number; METEOR is the grader’s own exact token variant, without stemming or a synonym table, averaged per case; and RadFact runs with negative filtering disabled, so absent findings count in both directions\. Both lexical metrics were reimplemented in pure Python and asserted equal to the reference to machine precision over randomised corpora; without that parity, tuning optimises a function other than the one being scored\.

#### The marginal rule for emitting a statement:

RadFact splits a report into verifiable phrases and asks an entailment model whether each is supported by the reference, and symmetrically whether each reference phrase is supported by the report\. Let a candidate carrynnphrases of whicheeare entailed, so precision isp=e/np=e/n\. Appending one phrase entailed with probabilityqqmoves expected precision to\(e\+q\)/\(n\+1\)\(e\+q\)/\(n\+1\), an improvement exactly whenq\>pq\>p, while recall moves only if the phrase covers a reference phrase not already covered\. The emission rule is thereforeq\>p−Δ​R​\(∂F1/∂R\)​\(∂F1/∂P\)−1q\>p\-\\Delta R\\,\(\\partial F\_\{1\}/\\partial R\)\(\\partial F\_\{1\}/\\partial P\)^\{\-1\}, a per statement threshold rather than a global topkkcut, sinceppandΔ​R\\Delta Rdiffer for every candidate\. The lexical half agrees\. METEOR combines its components asF=10​P​R/\(R\+9​P\)F=10PR/\(R\+9P\), recall weighted about nine to one, and its chunk penalty0\.5​\(chunks/matches\)30\.5\(\\text\{chunks\}/\\text\{matches\}\)^\{3\}rewards contiguous reuse of phrasing clinicians actually wrote\. Together these select an architecture: text is assembled by choosing among sentences observed in the training corpus rather than generated freely, so an entailment failure can only come from choosing the wrong finding, never from invented language\[[13](https://arxiv.org/html/2609.13238#bib.bib16)\]\.

#### An offline surrogate for the hidden term:

Fitting thresholds requires tens of thousands of full corpus evaluations, which rules out a language model judge in the loop\. A deterministic lexical surrogate is used for search instead, scoring phrase entailment from ontology concept agreement, polarity, laterality, tooth identifiers and token containment, with contradictory polarity and swapped laterality hard zeroed so it cannot reward a class of error the real judge exists to punish\. One corpus evaluation completes in 0\.17 seconds once entailment matrices are precomputed and METEOR alignment is replaced by anO⁡\(P\+R\)O\(P\+R\)matcher proven equal to the grader’s greedy assignment\. Over 250 held out cases it separates a report written for one patient from one written for another at an area under the curve of 0\.9872, at mean similarity 0\.707 against 0\.217\. It ranks decoder variants and is never reported here as the challenge metric\.

## 3Corpus and Label Space

Table[1](https://arxiv.org/html/2609.13238#S3.T1)records the release as measured\. The field of view is far smaller and flatter than a head CT, so an initial77×134×13477\\times 134\\times 134mm crop was roughly 70% air; refitting to56×96×9656\\times 96\\times 96mm at 0\.5 mm cut padding to 38%\. Crops centre on the dentition, isolated by a high intensity percentile, rather than on all bone, whose centroid drifts towards the skull on a large acquisition\. Spacing is preserved rather than normalised away, since the reports quantify physical distances such as bone height above the sinus floor, and 367 cases carry more than one report\.

Table 1:The public release as measured\. Extent is the physical field of view along each acquisition axis, median with the 5th and 95th percentiles\.Reports are dictated Italian rendered into English, and the translation leaves traces that must be handled explicitly\. Tooth identifiers appear bare, as in “33 abutment of a prosthetic bridge; 34 absent”, so extraction cannot require a keyword prefix, and the token*mm\.*ends sentences constantly, so guarding it as an abbreviation merges two findings into one phrase the metric then scores as a single indivisible unit\. The most consequential artefact was lexical: Italian*incluso*denotes an impacted tooth, so a naive ontology matched bare “included” as impaction and fired on every sentence describing what the acquisition contains\. That covered 16\.3% of reference phrases, 86\.4% of them field of view statements; requiring a tooth identifier beside the trigger reduced the concept to 4\.5%\. A second defect inverted polarity on 4\.4% of phrases, because the trigger for the absence concept is itself a negation cue\. Extending the vocabulary to the terms actually used raised concept coverage from 81% to 90%\.

Phrases are canonicalised, vectorised and clustered by agglomerative cosine linkage, and each cluster becomes one reportable statement represented by the member maximising expected score against the others\. Labels union the findings across every report belonging to a case, since a finding one clinician recorded is present in the scan whether or not the selected reference mentions it\. The label space was fixed by oracle score, what a perfect predictor would attain by emitting exactly the positively labelled statements, which upper bounds everything downstream at no training cost\. Masking tooth identifiers collapses every absence sentence into one cluster asserting a list of teeth wrong for almost every case: 507 statements at 84\.6% coverage and oracle 0\.5406, against 512 at 82\.1% and 0\.5645 tooth aware\. The delivered space keeps 989 tooth aware statements at 89\.0% coverage and oracle 0\.6008\.

## 4What the Image Predicts, and What It Does Not

A 29 million parameter encoder was trained on the release: multi planar slices through an ImageNet pretrained backbone\[[9](https://arxiv.org/html/2609.13238#bib.bib17),[21](https://arxiv.org/html/2609.13238#bib.bib11)\], pooled by gated attention multiple instance learning\[[12](https://arxiv.org/html/2609.13238#bib.bib10)\], under an asymmetric multi label loss\[[19](https://arxiv.org/html/2609.13238#bib.bib9)\]with the classifier bias initialised to the corpus log odds\. Training was clean, with loss falling from 97 to 60 across five patient grouped folds and validation mean average precision drifting from 0\.057 to 0\.072\.

0\.40\.40\.50\.50\.60\.60\.70\.70\.80\.80\.90\.9Osteolytic lesionCondyles not incl\.Mandibular bodyCanal regularMaxilla partialSinuses minimalCondyles excludedMandible incl\.Out of fold area under the curveGeometryIntensity\(a\)Per statement discriminationallprev\.n≥12n\\\!\\geq\\\!120\.450\.450\.50\.50\.550\.550\.60\.60\.650\.650\.70\.7Mean AUCPriorEncoderLinear\(b\)Predictor comparison
Figure 2:Out of fold discrimination\. \(a\) Nine header numbers separate every field of view statement and fail on pathology, correctly: a lesion is not a property of the acquisition\. \(b\) The encoder sits below the corpus prior at all three aggregations\.All of that was noise\. Mean average precision over roughly one thousand statements is dominated by columns seen two or three times, where it is essentially random, so the aggregate barely moves whatever the model does\. Per statement area under the curve against the corpus prior, which scores exactly 0\.500 by construction, answers directly: every support band sits at or below chance, the prevalence weighted figure is 0\.486 over 985 evaluable statements, and for each prevalent statement the mean predicted probability was the same for positives and negatives\. The model had memorised 497 cases and generalised nothing\.

The decisive follow up was a control rather than a larger model\. Nine numbers read from the image header alone, the logarithm of the volume dimensions, the voxel spacing and the logarithm of the physical extent, were fitted with one cross validated logistic regression per statement\. Field of view statements, the most frequent in the corpus, are strongly predicted \(Figure[2](https://arxiv.org/html/2609.13238#S4.F2)a\)\. Over the 37 statements with adequate support, header geometry averages 0\.627 against 0\.542 for intensity statistics computed after resampling, and adding intensity moves the mean only to 0\.637 while requiring pixel decode, so the header only path was kept\. The task is learnable from information the network already received on its auxiliary input, so the deep model was the problem rather than the problem being impossible\. One strongly regularised linear model per statement over a 122 dimensional global descriptor\[[17](https://arxiv.org/html/2609.13238#bib.bib13)\]raises prevalence weighted discrimination to 0\.593 and reaches 0\.669 where support allows \(Figure[2](https://arxiv.org/html/2609.13238#S4.F2)b\), in a deployable form of 52 kB against 580 MB of fold checkpoints\.

### 4\.1Why No Trained Model Wins the Lexical Column

A sharper diagnosis explains that result and bounds what any model on this release can achieve\. Predicting which prototype sentence a report uses from*acquisition centre alone*reaches a mean area under the curve of 0\.718 over the 60 most frequent sentences, against 0\.663 for the image derived model, and 72\.2% of a sentence’s usage falls inside its single most used centre\. Sentence choice is house style, not anatomy: the encoder was not failing to see the jaw, it was being asked to predict a dictation convention, which is not in the pixels\.

Every subsequent observation follows\. A perfect per case selector reaches BLEU\-4 0\.4086 against 0\.1663 for a single fixed report, but that gap is largely the reward for reproducing one centre’s phrasing verbatim, and centre P supplies 412 of 622 cases\. The evaluation centre has conventions absent from all four training centres, so the gap is unreachable there: the observed test BLEU\-4 of 0\.0943 falls below every leave one centre out fold, which range from 0\.1241 to 0\.1968\. Perfect knowledge of which teeth are absent, the most visually obvious content in the volume, is worth 0\.0041 BLEU\-4, since the n\-grams live in the sentence frames rather than the identifiers, and clustering cases by geometry to fit one report per cluster loses at every cluster count, by 0\.0225 to 0\.0314 BLEU\-4\.

## 5Constrained Decoding

Selection is greedy coordinate ascent against Equation[1](https://arxiv.org/html/2609.13238#S2.E1)on out of fold probabilities, warm started from the best threshold vector found for any predictor, and converges in six rounds from 0\.3001 to 0\.3536\. Three constraints are then imposed, each derived from the arithmetic of Section[2](https://arxiv.org/html/2609.13238#S2)rather than from stylistic preference\.

#### Contradiction:

The metric scores statements one at a time and never reads the report as a whole, so asserting both sides of a contradiction is profitable: each side earns credit on the subset of cases whose reference agrees with it\. Unconstrained selection duly produced a report in which the maxilla was both absent from and partially included in the acquisition, and teeth 38 and 48 were simultaneously missing, semi impacted and erupted in the arch\. Such a report is guaranteed to be wrong about something, so the constraint is not a cost paid for readability: removing one misencoded absence statement raised the composite score from 0\.4023 to 0\.4061\. Three rules apply, namely opposite polarity on a shared concept, incompatible values from one attribute vocabulary such as lingual against buccal, and a tooth asserted absent in one statement and described in another\.

#### Redundancy and specificity:

Two statements carrying the same concept set, laterality and tooth identifiers assert one fact\. Under lexical scoring they are two separate wins, since each matches a different phrasing somewhere in the corpus, and selection duly reached for seven paraphrases of a lingual canal course\. Under the entailment metric they are one fact asserted twice: recall cannot move, since the phrase they entail is entailed already, and precision is neutral at best and halved on that fact when it is false\. Selection therefore refuses a statement whose assertion signature is present already\. Specificity follows the same accounting: statements naming particular teeth are 48% of the pool, yet their mean prevalence is 0\.016 against 0\.035 for statements naming none, so in a report served to every patient a tooth identifier is wrong about 98% of the time\. Banning them wholesale cost 0\.079 composite score; correcting the two encoding defects that let the contradictions through recovered it instead\.

### 5\.1Report Length and Geometry Conditioning

Left unconstrained the ascent grows the report to 53 statements at composite score 0\.4882 and continues improving\. The objective genuinely prefers that, and it is the wrong report to deliver, since references average 8\.9 verifiable phrases and the automatic score is a filter in front of a blinded clinical comparison rather than the prize itself\. Marginal value collapses long before the optimum, falling from 0\.0070 composite score per added statement between eleven and fifteen statements to 0\.0043, 0\.0033 and finally 0\.0007 at twenty, twenty six and fifty three, so length is taken from the knee rather than from the argmax\.

The delivered decoder keeps eight statements unconditional and converts five into logistic gates over the nine header features, emitting 11\.2 sentences on average against 13\.0 for the constant report, with thresholds fitted on out of fold gate probabilities and kept only when they raise the composite score\. Gating pays twice: precision divides by the number of statements emitted, so not paying for a false statement is a direct gain, and the final ranking includes a blinded pairwise comparison by clinicians, where a report identical for every patient is the first thing a reader notices\. It is restricted to the existing core, since allowing additions took a fifteen statement core to a mean of 20\.9 sentences; that restriction also makes consistency free, every per case report being a subset of a core already checked for contradiction and redundancy\.

## 6Results

Table[2](https://arxiv.org/html/2609.13238#S6.T2)scores every decoder variant over the 622 public cases, with clinical figures from the surrogate of Section[2](https://arxiv.org/html/2609.13238#S2)\. The ordering is the central result: the variant with the highest BLEU\-4 has the lowest composite score, and the gap is not marginal, since selecting against the visible projection costs 0\.121 composite score as entailment precision falls from 0\.522 to 0\.266 \(Figure[3](https://arxiv.org/html/2609.13238#S6.F3)\)\. An exhaustive search from that report found no addition, deletion or substitution raising the composite score at fixed BLEU\-4 and METEOR, so the two objectives are opposed there rather than untuned\.

Table 2:Decoder variants over 622 public cases, out of fold\. Precision and recall come from the offline surrogate and are not the challenge metric\.0\.110\.110\.120\.120\.130\.130\.140\.140\.150\.150\.160\.160\.170\.170\.180\.180\.280\.280\.320\.320\.360\.360\.40\.4lexically tuneddeliveredencoder 29Mlinear 122corpus priorobjective tuned−0\.040\-0\.040BLEU\-4\+0\.121\+0\.121compositeBLEU\-4, the visible projectionComposite scoreFigure 3:The visible signal is anticorrelated with the objective it projects: of the systems considered, the one with the best BLEU\-4 scores worst on the quantity that decides the outcome\.#### Held out centre and transfer:

The evaluation set contains 50 cases from a centre absent from training, so stratified fold numbers measure interpolation and read higher than the outcome\. On that set the delivered entry measured BLEU\-40\.0943±0\.03540\.0943\\pm 0\.0354and METEOR0\.3542±0\.05990\.3542\\pm 0\.0599, placing first on METEOR and third overall on the aggregated automatic ranking\. Two transfer measurements follow and are reported separately rather than folded into one: against an in sample pooled fit, BLEU\-4 arrived at 0\.567 of its development value and METEOR at 0\.987, while against a leave one centre out estimate the same result corresponds to 0\.66 and 1\.015\. Exact 4\-gram sequences do not survive an unseen dictation style, and token level overlap does, which is the signature predicted in Section[4](https://arxiv.org/html/2609.13238#S4)and why the delivered entry weights centres equally rather than fitting the pooled corpus\. The BLEU\-4 standing is the anticipated cost of the trade in Figure[3](https://arxiv.org/html/2609.13238#S6.F3): 4\-gram overlap is bound most tightly to centre specific phrasing and carries least weight in Equation[1](https://arxiv.org/html/2609.13238#S2.E1)\.

#### Clinical Inference and Safety Implications:

The experiments separate two forms of report content that have different clinical interpretations\. Statements describing acquisition coverage, such as inclusion of the mandible, condyles or maxilla, are recoverable from header geometry and can therefore be conditioned on each examination\. In contrast, the present experiments do not establish reliable image\-based discrimination for low\-prevalence pathological or tooth\-specific findings\. The generated report should therefore be interpreted as a constrained, high\-precision summary of findings supported by the benchmark signal rather than as a comprehensive diagnostic reading of the CBCT\[[10](https://arxiv.org/html/2609.13238#bib.bib20)\]\.

The distinction provides the necessary flexibility and caution for clinical deployment\. A sentence selected because it is predictable from field\-of\-view geometry indicates what anatomy is available for assessment, not whether that anatomy is normal\. Conversely, absence of a pathology statement from the generated report cannot be interpreted as evidence that the pathology is absent\[[3](https://arxiv.org/html/2609.13238#bib.bib22)\]\. The contradiction, laterality and tooth\-consistency constraints reduce internally impossible assertions, but they do not substitute for direct localisation or clinician verification\. Consequently, the system is best positioned as a report\-drafting or benchmark inference tool whose patient\-specific outputs require review before clinical use\[[1](https://arxiv.org/html/2609.13238#bib.bib21)\]\.

## 7Deployment

An earlier submission returned BLEU\-40\.0161±0\.01290\.0161\\pm 0\.0129and METEOR0\.1088±0\.02840\.1088\\pm 0\.0284\. The diagnosis was arithmetic: the hardcoded 52 token fallback paragraph scores0\.0111±0\.01390\.0111\\pm 0\.0139and0\.0984±0\.03500\.0984\\pm 0\.0350against training references, matching mean and spread, so all 50 cases had received it\. The platform mounts its own volume over the model directory, shadowing whatever the image baked there, which nothing local reproduces: the container ran, wrote a report and exited zero with the model intact\.

Four consequences follow\. The report and its gate arithmetic ship as an importable module rather than a data file, since a mount cannot shadow an import\. Inference reads only the image header, so no decode failure, memory limit or missing accelerator can break it\. Four independent layers each write a valid report alone: gated selection, the constant report, an embedded fallback and a bare write\. And the container test mounts an empty directory over both model directories, asserting the report is byte identical to the unshadowed run rather than merely present, which is the check the original test missed\. Nine scenarios are exercised, among them corrupt input, read only root, capped memory and two geometries that must differ\.

## 8Discussion and Conclusion

The recurring lesson concerns instrument selection: mean average precision drifted upward while the model was getting no better, and a thirty second control over nine header numbers was worth more than the accelerator hour it audited\. The same holds one level up: a leaderboard exposing one fifth of the scoring function is an instrument, and calibrating against it blind produces confident movement in the wrong direction\[[14](https://arxiv.org/html/2609.13238#bib.bib7),[15](https://arxiv.org/html/2609.13238#bib.bib8)\]\. Three limitations bound the conclusions\. Clinical figures are surrogate estimates, since the judge carrying 80% of the weight was unavailable during development; the house style result bounds what is reachable on this release, not in principle; and the decoder conditions on geometry rather than anatomy, so it is silent about pathology, as Figure[2](https://arxiv.org/html/2609.13238#S4.F2)a shows for the lesion statement\. Closing that gap calls for the localisation supervision the preceding segmentation releases provide\[[7](https://arxiv.org/html/2609.13238#bib.bib3),[2](https://arxiv.org/html/2609.13238#bib.bib19)\]\. Building against a composite objective whose dominant term is hidden during development is a reasoning problem before it is a modelling problem: reproducing the grader exactly, measuring which signals are recoverable, and deriving decoding constraints from the scored metric produced a system scoring 0\.4122 against 0\.2909 for a lexically tuned counterpart\.

## References

- \[1\]G\. Arun, R\. Syam, A\.A\. Nair,et al\.\(2025\)An integrated framework for ethical healthcare chatbots using LangChain and NeMo guardrails\.AI Ethics5,pp\. 3981–3992\.Cited by:[§6](https://arxiv.org/html/2609.13238#S6.SS0.SSSx2.p2.1)\.
- \[2\]G\. Arun, M\.G\. Vazhunnavar, A\.B\. George,et al\.\(2026\)nnU\-Net based automated nasopalatine nerve segmentation for maxillofacial robotic surgery\.Journal of Maxillofacial and Oral Surgery\.Cited by:[§8](https://arxiv.org/html/2609.13238#S8.p1.1)\.
- \[3\]G\. Arun, S\. Bathini, and B\. BhasuranTemporal epistemic graph of thought: uncertainty over latent reasoning structure for selective abstention in clinical prediction\.In5th Workshop on NLP for Positive Impact,Cited by:[§6](https://arxiv.org/html/2609.13238#S6.SS0.SSSx2.p2.1)\.
- \[4\]S\. Bannur, K\. Bouzid, D\. C\. Castro, A\. Schwaighofer, A\. Thieme, S\. Bond\-Taylor, M\. Ilse, F\. Pérez\-García, V\. Salvatelli, H\. Sharma, F\. Meissen, M\. Ranjit, S\. Srivastav, J\. Gong, N\. C\. F\. Codella, F\. Falck, O\. Oktay, M\. P\. Lungren, M\. T\. Wetscherek, J\. Alvarez\-Valle, and S\. L\. Hyland\(2024\)MAIRA\-2: grounded radiology report generation\.arXiv preprint arXiv:2406\.04449\.External Links:[Document](https://dx.doi.org/10.48550/arXiv.2406.04449)Cited by:[§1](https://arxiv.org/html/2609.13238#S1.p2.1)\.
- \[5\]S\. Bird, E\. Klein, and E\. Loper\(2009\)Natural language processing with Python\.O’Reilly Media\.Cited by:[§2](https://arxiv.org/html/2609.13238#S2.p1.2)\.
- \[6\]F\. Bolelli, A\. Ben\-Hamadou, L\. Lumetti, S\. Pujades Rocamora, N\. van Nistelrooij, K\. Marchesini, F\. Cremonini, M\. Di Bartolomeo, L\. Fix, N\. Morelli, A\. Rekik, N\. Neifar, O\. Abida, O\. Smaoui, T\. Xi, S\. Vinayahalingam, L\. Lombardo, A\. Anesi, and C\. Grana\(2026\)ODIN2026 challenge: multimodal text report generation for oral and dental image analysis\.Zenodo\.External Links:[Document](https://dx.doi.org/10.5281/zenodo.19727377)Cited by:[§1](https://arxiv.org/html/2609.13238#S1.p1.1),[§2](https://arxiv.org/html/2609.13238#S2.p1.2)\.
- \[7\]F\. Bolelli, L\. Lumetti, N\. van Nistelrooij, S\. Vinayahalingam, M\. Di Bartolomeo, K\. Marchesini, A\. Pellacani, E\. Candeloro, G\. Rosati, T\. Xi, F\. Isensee, Y\. Kirchhoff, L\. Krämer, M\. Rokuss, C\. Ulrich, K\. Maier\-Hein, R\. Grosu, S\. Bergé, A\. Anesi, and C\. Grana\(2026\)Multi\-structure segmentation in CBCT volumes: the ToothFairy2 challenge\.Medical Image Analysis112,pp\. 104095\.External Links:[Document](https://dx.doi.org/10.1016/j.media.2026.104095)Cited by:[§1](https://arxiv.org/html/2609.13238#S1.p1.1),[§8](https://arxiv.org/html/2609.13238#S8.p1.1)\.
- \[8\]F\. Bolelli, L\. Lumetti, S\. Vinayahalingam, M\. Di Bartolomeo, A\. Pellacani, K\. Marchesini, N\. van Nistelrooij, P\. van Lierop, T\. Xi, Y\. Liu, R\. Xin, T\. Yang, L\. Wang, C\. Wu, D\. Xu, Z\. Cui, S\. Chen, Z\. Wang, A\. Anesi, S\. Bergé, and C\. Grana\(2025\)Segmenting the inferior alveolar canal in CBCT volumes: the ToothFairy challenge\.IEEE Transactions on Medical Imaging44\(4\),pp\. 1890–1906\.External Links:[Document](https://dx.doi.org/10.1109/TMI.2024.3523096)Cited by:[§1](https://arxiv.org/html/2609.13238#S1.p1.1)\.
- \[9\]J\. Deng, W\. Dong, R\. Socher, L\. Li, K\. Li, and L\. Fei\-Fei\(2009\)ImageNet: a large\-scale hierarchical image database\.InIEEE Conference on Computer Vision and Pattern Recognition \(CVPR\),pp\. 248–255\.External Links:[Document](https://dx.doi.org/10.1109/CVPR.2009.5206848)Cited by:[§4](https://arxiv.org/html/2609.13238#S4.p1.1)\.
- \[10\]A\.B\. George, S\. Bathini, and G\.A\. Govind\(2025\)Grad\-CAM and Grad\-CAM\+\+ for explainable oral squamous cell carcinoma detection on orthopantomograms\.In2025 SENNET,Vellore, India,pp\. 1–6\.Cited by:[§6](https://arxiv.org/html/2609.13238#S6.SS0.SSSx2.p1.1)\.
- \[11\]A\.B\. George, P\. Gadha Lekshmi, S\. Bathini, and A\. Govind\(2026\)Semi\-supervised CBCT–IOS registration using PointNetLK\.InOral and Dental Image Analysis,LNCS, Vol\.16473,Cham\.Cited by:[§1](https://arxiv.org/html/2609.13238#S1.p1.1)\.
- \[12\]M\. Ilse, J\. M\. Tomczak, and M\. Welling\(2018\)Attention\-based deep multiple instance learning\.InProceedings of the 35th International Conference on Machine Learning \(ICML\),pp\. 2127–2136\.Cited by:[§4](https://arxiv.org/html/2609.13238#S4.p1.1)\.
- \[13\]Z\. Ji, N\. Lee, R\. Frieske, T\. Yu, D\. Su, Y\. Xu, E\. Ishii, Y\. J\. Bang, A\. Madotto, and P\. Fung\(2023\)Survey of hallucination in natural language generation\.ACM Computing Surveys55\(12\),pp\. 1–38\.External Links:[Document](https://dx.doi.org/10.1145/3571730)Cited by:[§2](https://arxiv.org/html/2609.13238#S2.SS0.SSSx1.p1.1)\.
- \[14\]L\. Maier\-Hein, M\. Eisenmann, A\. Reinke, S\. Onogur, M\. Stankovic, P\. Scholz,et al\.\(2018\)Why rankings of biomedical image analysis competitions should be interpreted with care\.Nature Communications9,pp\. 5217\.External Links:[Document](https://dx.doi.org/10.1038/s41467-018-07619-7)Cited by:[§8](https://arxiv.org/html/2609.13238#S8.p1.1)\.
- \[15\]L\. Maier\-Hein, A\. Reinke, M\. Kozubek, A\. L\. Martel, T\. Arbel, M\. Eisenmann,et al\.\(2020\)BIAS: transparent reporting of biomedical image analysis challenges\.Medical Image Analysis66,pp\. 101796\.External Links:[Document](https://dx.doi.org/10.1016/j.media.2020.101796)Cited by:[§8](https://arxiv.org/html/2609.13238#S8.p1.1)\.
- \[16\]K\. Papineni, S\. Roukos, T\. Ward, and W\. Zhu\(2002\)BLEU: a method for automatic evaluation of machine translation\.InProceedings of the 40th Annual Meeting of the Association for Computational Linguistics,pp\. 311–318\.External Links:[Document](https://dx.doi.org/10.3115/1073083.1073135)Cited by:[§2](https://arxiv.org/html/2609.13238#S2.p1.2)\.
- \[17\]F\. Pedregosa, G\. Varoquaux, A\. Gramfort, V\. Michel, B\. Thirion, O\. Grisel,et al\.\(2011\)Scikit\-learn: machine learning in Python\.Journal of Machine Learning Research12,pp\. 2825–2830\.Cited by:[§4](https://arxiv.org/html/2609.13238#S4.p3.1)\.
- \[18\]E\. Reiter\(2018\)A structured review of the validity of BLEU\.Computational Linguistics44\(3\),pp\. 393–401\.External Links:[Document](https://dx.doi.org/10.1162/coli%5Fa%5F00322)Cited by:[§1](https://arxiv.org/html/2609.13238#S1.p2.1)\.
- \[19\]T\. Ridnik, E\. Ben\-Baruch, N\. Zamir, A\. Noy, I\. Friedman, M\. Protter, and L\. Zelnik\-Manor\(2021\)Asymmetric loss for multi\-label classification\.InProceedings of the IEEE/CVF International Conference on Computer Vision \(ICCV\),pp\. 82–91\.External Links:[Document](https://dx.doi.org/10.1109/ICCV48922.2021.00015)Cited by:[§4](https://arxiv.org/html/2609.13238#S4.p1.1)\.
- \[20\]X\. Wang, H\. Ouyang, B\. Bhasuran, X\. Luo, K\. Hanna, M\. L\. A\. Lustria, and Z\. He\(2024\)Lab\-ai–retrieval\-augmented language model for personalized lab test interpretation in clinical medicine\.arXiv e\-prints\.Cited by:[§1](https://arxiv.org/html/2609.13238#S1.p1.1)\.
- \[21\]R\. Wightman\(2019\)PyTorch image models\.Note:[https://github\.com/huggingface/pytorch\-image\-models](https://github.com/huggingface/pytorch-image-models)External Links:[Document](https://dx.doi.org/10.5281/zenodo.4414861)Cited by:[§4](https://arxiv.org/html/2609.13238#S4.p1.1)\.
- \[22\]F\. Yu, M\. Endo, R\. Krishnan, I\. Pan, A\. Tsai, E\. P\. Reis, E\. K\. U\. N\. Fonseca, H\. M\. H\. Lee, Z\. S\. H\. Abad, A\. Y\. Ng, C\. P\. Langlotz, V\. K\. Venugopal, and P\. Rajpurkar\(2023\)Evaluating progress in automatic chest X\-ray radiology report generation\.Patterns4\(9\),pp\. 100802\.External Links:[Document](https://dx.doi.org/10.1016/j.patter.2023.100802)Cited by:[§1](https://arxiv.org/html/2609.13238#S1.p2.1)\.