Partial Information Decomposition as a Multi-Contrast 3D MRI Selection Strategy for Resource-Constrained Deep Neural Network Training in Brain Tumor Segmentation

# Partial Information Decomposition as a Multi-Contrast 3D MRI Selection Strategy for Resource-Constrained Deep Neural Network Training in Brain Tumor Segmentation
Source: [https://arxiv.org/html/2607.15396](https://arxiv.org/html/2607.15396)
Agamdeep S\. ChopraORCID:[0009\-0003\-4386\-7894](https://orcid.org/0009-0003-4386-7894)Department of Mechanical Engineering, University of Washington, Seattle, WA, USAMehmet KurtORCID:[0000\-0002\-5618\-0296](https://orcid.org/0000-0002-5618-0296)Department of Mechanical Engineering, University of Washington, Seattle, WA, USA

###### Abstract

Multi\-contrast 3D MRI segmentation can be computationally demanding when all available sequences are used\. We evaluate a pre\-training Partial Information Decomposition framework that ranks input pairs according to their redundant, unique, and synergistic information about regional tumor burden and selects the highest\-ranked pair for downstream training\. Applied to T1n, T1c, T2w, and T2\-FLAIR MRI, the framework selected T1c\+T2\-FLAIR\. We then trained eleven architecturally identical lightweight 3D U\-Nets using different input configurations\. On an independent test cohort, T1c\+T2\-FLAIR was the strongest two\-input configuration and ranked second overall in mean Dice \(0\.676 versus 0\.687 for all four inputs\)\. Independent Shapley analysis on the full\-input model also identified T2\-FLAIR and T1c as the most influential inputs and their pairwise interaction as the strongest\. These findings demonstrate the practical value of PID based pre\-training selection for identifying compact, informative MRI input sets before costly 3D model development\.

Keywords:Partial information decomposition, MRI sequence selection, Brain tumor segmentation, Resource\-constrained learning

## 1Introduction

Multi\-contrast MRI supports brain tumor segmentation because each sequence emphasizes complementary tissue characteristics\. T1\-weighted MRI depicts anatomy, contrast enhanced T1 highlights enhancing tumor, T2 MRI is sensitive to fluid, and T2\-FLAIR suppresses cerebrospinal fluid to emphasize edema and infiltrative abnormalities\. Consequently, brain tumor benchmarks and high\-performing 3D segmentation frameworks commonly integrate all available contrasts for multi\-region segmentation\[[13](https://arxiv.org/html/2607.15396#bib.bib1),[6](https://arxiv.org/html/2607.15396#bib.bib2),[4](https://arxiv.org/html/2607.15396#bib.bib4),[9](https://arxiv.org/html/2607.15396#bib.bib3)\]\.

Although, this approach can be expensive in computationally constrained settings\. Each additional sequence increases storage, preprocessing, data\-transfer, memory, and training costs, which are amplified in 3D by large patch and volume tensors and by repeated experiments across folds, architectures, and hyperparameter settings\. The practical question is therefore whether a smaller input subset can be selected before substantial resources are committed to downstream model development\.

Standard subset selection relies on either exhaustive model training or single\-source relevance measures\. Exhaustive evaluation provides direct task performance estimates but scales poorly and must be repeated for each downstream architecture\. Mutual information based methods offer a cheaper alternative by balancing target relevance against redundancy\[[14](https://arxiv.org/html/2607.15396#bib.bib5)\], but they may overlook synergy, whereby two sequences are jointly informative even when neither is individually dominant\.

![Refer to caption](https://arxiv.org/html/2607.15396v1/figures/PID.png)Figure 1:Diagram representing partial information decomposition for two sources,X1X\_\{1\}andX2X\_\{2\}, relative to a targetYY\. The total information is partitioned into source\-specific contributions,U1U\_\{1\}andU2U\_\{2\}, redundant informationRRprovided by both sources, and synergistic informationSSavailable only from their joint observation\.Partial information decomposition \(PID\) separates target information into redundant, unique, and synergistic components\[[19](https://arxiv.org/html/2607.15396#bib.bib6),[2](https://arxiv.org/html/2607.15396#bib.bib7)\]\. It has been used to define feature relevance and guide feature selection\[[20](https://arxiv.org/html/2607.15396#bib.bib8),[18](https://arxiv.org/html/2607.15396#bib.bib9)\], while recent works such as SFL\-Net applied PID\-guided latent factorization to multi\-contrast MRI to PET synthesis\[[3](https://arxiv.org/html/2607.15396#bib.bib12)\]\. Building on this work, prior studies of MRI sequence contributions in brain tumor segmentation\[[21](https://arxiv.org/html/2607.15396#bib.bib10),[16](https://arxiv.org/html/2607.15396#bib.bib11)\], and contrast\-level Shapley attribution\[[15](https://arxiv.org/html/2607.15396#bib.bib19)\], we introduce a generalnn\-to\-2 pre\-training framework for multi\-contrast MRI tumor segmentation\. A compact autoencoder is trained for each input sequence, the resulting embeddings are quantized, and all\(n2\)\\binom\{n\}\{2\}pairs are ranked using a region\-aware minimum mutual information PID \(MMI\-PID\) definition score\[[1](https://arxiv.org/html/2607.15396#bib.bib21)\]\. The highest scoring pair is then selected for downstream model development\. We evaluate the framework using four CoRe\-BT MRI sequences\[[6](https://arxiv.org/html/2607.15396#bib.bib2)\], comparing the full four\-input model, all six two\-input combinations, and four single\-input models with patient\-level statistical testing and Shapley attribution under a restricted compute budget\.

## 2Method

### 2\.1PID Pair Selection

![Refer to caption](https://arxiv.org/html/2607.15396v1/figures/pipeline.drawio.png)Figure 2:PID pair\-selection pipeline\. Spatially aligned patches from each MRI sequence are encoded using sequence\-specific autoencoders\. The frozen latent representations are quantized into discrete source codes, while aligned tumor masks provide discretized regional tumor\-burden targets\. All two\-input combinations are evaluated using region\-aware PID, ranked by their aggregate scores, and the highest\-ranked pair is retained for downstream model training\.Figure[2](https://arxiv.org/html/2607.15396#S2.F2)summarizes the selection process\. For each training subject, we sampled 128 spatially corresponding32×32×3232\\times 32\\times 32patches from all MRI sequences and the tumor mask\. 70% of patch centers were drawn from the tumor foreground\. For each tumor regionc∈\{edema,enhancing tumor,necrotic\}c\\in\\\{\\text\{edema\},\\text\{enhancing tumor\},\\text\{necrotic\}\\\}, tumor burden was defined as the fraction of patch voxels assigned to that region\. Zero\-burden patches formed class 0, while positive burdens were divided into three training\-set quantile classes\.

A separate shallow 3D autoencoder was trained for each MRI sequence to reconstruct its input patches using mean\-squared error\. The encoder comprised three3×3×33\\times 3\\times 3convolutional blocks with stride two and 8, 16, and 32 output channels respectively\. Each convolution was followed by instance normalization and a ReLU activation\. For a32×32×3232\\times 32\\times 32input patch, the encoder produced a32×4×4×432\\times 4\\times 4\\times 4feature map, which was flattened and linearly projected to a 32\-dimensional latent vector\. After training, the encoders were frozen and the decoders were discarded\. For each sequencemm, a separateKK\-means model withK=16K=16was fitted to the training latent vectors, and each cluster assignment defined the discrete source codeqmq\_\{m\}\[[7](https://arxiv.org/html/2607.15396#bib.bib13),[10](https://arxiv.org/html/2607.15396#bib.bib14)\]\.

For every input pair\(a,b\)\(a,b\)and tumor regioncc, mutual information was estimated from the empirical joint distribution of the two source codes and the discretized tumor\-burden targetTcT\_\{c\}\[[5](https://arxiv.org/html/2607.15396#bib.bib15)\]\. Following the bivariate PID framework\[[19](https://arxiv.org/html/2607.15396#bib.bib6)\], we decomposed the information that each pair of quantized source codes provided about the regional tumor\-burden target\. We used the MMI formulation as a computationally simple proxy for bivariate PID\[[1](https://arxiv.org/html/2607.15396#bib.bib21)\]\. For input pair\(a,b\)\(a,b\)and tumor regioncc, the PID terms were

Ra​bc\\displaystyle R\_\{ab\}^\{c\}=min⁡\{I​\(qa;Tc\),I​\(qb;Tc\)\},\\displaystyle=\\min\\\!\\left\\\{I\(q\_\{a\};T\_\{c\}\),I\(q\_\{b\};T\_\{c\}\)\\right\\\},Ua\|bc\\displaystyle U\_\{a\|b\}^\{c\}=I​\(qa;Tc\)−Ra​bc,Ub\|ac=I​\(qb;Tc\)−Ra​bc,\\displaystyle=I\(q\_\{a\};T\_\{c\}\)\-R\_\{ab\}^\{c\},\\qquad U\_\{b\|a\}^\{c\}=I\(q\_\{b\};T\_\{c\}\)\-R\_\{ab\}^\{c\},Sa​bc\\displaystyle S\_\{ab\}^\{c\}=I​\(\(qa,qb\);Tc\)−Ra​bc−Ua\|bc−Ub\|ac\.\\displaystyle=I\(\(q\_\{a\},q\_\{b\}\);T\_\{c\}\)\-R\_\{ab\}^\{c\}\-U\_\{a\|b\}^\{c\}\-U\_\{b\|a\}^\{c\}\.\(1\)
To rank the candidate pairs, we defined an aggregate score inspired by PID\-based feature selection criteria that favor informative and non\-redundant feature sets\[[20](https://arxiv.org/html/2607.15396#bib.bib8),[18](https://arxiv.org/html/2607.15396#bib.bib9)\]:

Score​\(a,b\)=∑cwc​\[I​\(\(qa,qb\);Tc\)\+λ​\(Ua\|bc\+Ub\|ac\+Sa​bc\)\],λ=0\.5,\\mathrm\{Score\}\(a,b\)=\\sum\_\{c\}w\_\{c\}\\left\[I\(\(q\_\{a\},q\_\{b\}\);T\_\{c\}\)\+\\lambda\\left\(U\_\{a\|b\}^\{c\}\+U\_\{b\|a\}^\{c\}\+S\_\{ab\}^\{c\}\\right\)\\right\],\\qquad\\lambda=0\.5,\(2\)wherewcw\_\{c\}was the normalized inverse frequency of positive patches for regioncc\. Thus, joint target information determined the baseline pair score, while unique and synergistic information received an additional weight relative to redundant information\. The selected pair was

ℳ∗=arg⁡max\{a,b\}⊂ℳ⁡Score​\(a,b\),\|ℳ∗\|=2\.\\mathcal\{M\}^\{\*\}=\\arg\\max\_\{\\\{a,b\\\}\\subset\\mathcal\{M\}\}\\mathrm\{Score\}\(a,b\),\\qquad\|\\mathcal\{M\}^\{\*\}\|=2\.\(3\)

### 2\.2Lightweight 3D U\-Net

The downstream segmentation model was a lightweight 3D U\-Net with three resolution levels\[[4](https://arxiv.org/html/2607.15396#bib.bib4)\]\. Each encoder and decoder block contained two3×3×33\\times 3\\times 3convolutions followed by ReLU activations and Instance Normalization\. Max pooling was used for downsampling, while nearest\-neighbor interpolation followed by1×1×11\\times 1\\times 1convolution was used for upsampling\. A final1×1×11\\times 1\\times 1convolution produced the output logits\.

Models were trained using64×64×6464\\times 64\\times 64patches, a batch size of one, and 256 patches per epoch\. The training objective combined Dice loss and binary cross\-entropy, balancing region overlap with voxelwise supervision\. Optimization used AdamW with a weight decay of10−410^\{\-4\}\[[12](https://arxiv.org/html/2607.15396#bib.bib16)\]\. The learning rate was reduced from10−310^\{\-3\}to10−510^\{\-5\}using a cosine schedule over 200 epochs\[[11](https://arxiv.org/html/2607.15396#bib.bib17)\]\. A 30 subject validation cohort was used for development stage benchmarking, and final evaluation was performed on an independent held\-out 28 subject test cohort\.

## 3Experiments

We used 132 training, 30 validation, and 28 test subjects from the MRI and tumor\-mask component of CoRe\-BT dataset\[[6](https://arxiv.org/html/2607.15396#bib.bib2)\]\. Subjects missing any required sequence were excluded\. Inputs were T1n, T1c, T2w, and T2\-FLAIR\. Reported targets were edema, enhancing tumor, and necrotic/non\-enhancing tumor\. Volumes were cropped, clipped to the 0\.5th–99\.5th percentile range, and normalized for training\.

PID selected T1c\+T2\-FLAIR as the highest\-ranked input pair for this task\. We then trained 11 architecturally identical 3D U\-Net models \(one four\-input model, all six pairwise models, and four single\-input models\)\. This design tested whether the PID ranking predicted downstream segmentation performance\.

The models were evaluated on the held\-out set using patient\-level macro\-averaged Dice, HD95, sensitivity, and precision across tumor regions\. Statistical analyses included 10,000 bootstrap resamples, Spearman correlation, paired Wilcoxon tests, Friedman tests with Holm correction, and a−0\.03\-0\.03Dice non\-inferiority margin\[[8](https://arxiv.org/html/2607.15396#bib.bib20)\]\. Shapley values over all 16 input coalitions were computed for the four\-input model, replacing omitted inputs with training set channel means inside the brain mask and zeros outside\[[17](https://arxiv.org/html/2607.15396#bib.bib18),[15](https://arxiv.org/html/2607.15396#bib.bib19)\]\.

The complete experiment used one RTX 3090 with GPU memory capped at 2 GB, four logical CPU cores, and 4 GB of system memory for training and feature extraction\.

## 4Analysis and Results

### 4\.1PID Selection and Segmentation

PID ranked T1c\+T2\-FLAIR first with a score of 0\.963\. Across the six pair models, the PID score was positively associated with test\-set Dice \(ρ=0\.600\\rho=0\.600,p=0\.208p=0\.208\) and negative HD95 \(ρ=0\.600\\rho=0\.600,p=0\.208p=0\.208\)\. Both associations were directionally consistent with downstream performance, but neither was statistically significant, and the analysis had limited power because only six pairs were available\.

Table 1:Region\-specific PID scores for all input\-pair combinations\. Nec\., Ed\., and Enh\. denote necrotic tumor, edema, and enhancing tumor, respectively\.On the test cohort, the full model achieved average Dice 0\.687 and HD95 16\.09\. T1c\+T2\-FLAIR achieved Dice 0\.676 and HD95 16\.77, making it the strongest pair and second overall \(Table[2](https://arxiv.org/html/2607.15396#S4.T2)\)\. The selected pair retained 98\.5% of the full\-model mean Dice while using half the input channels\.

Table 2:Test\-set performance for the top input configurations\. Ed, En, and Ne denote edema, enhancing tumor, and necrotic tumor, respectively\. HD95, average Dice, sensitivity, and precision are patient\-level macro averages across the three regions\.The selected vs\. full Dice difference was−0\.010\-0\.010\(95% CI\[−0\.052,0\.034\]\[\-0\.052,0\.034\]; paired Wilcoxonp=0\.0505p=0\.0505\)\. HD95 increased by 0\.68 \(95% CI\[−4\.54,5\.09\]\[\-4\.54,5\.09\];p=0\.174p=0\.174\)\. Sensitivity was nearly identical \(difference<0\.001<0\.001;p=0\.991p=0\.991\), and precision differed by−0\.015\-0\.015\(p=0\.099p=0\.099\)\. Dice non\-inferiority was not established given our choice of margin\.

Model configuration affected Dice \(χ102=164\.63\\chi^\{2\}\_\{10\}=164\.63,p<10−29p<10^\{\-29\}, Kendall’sW=0\.588W=0\.588\) and HD95 \(χ102=79\.07\\chi^\{2\}\_\{10\}=79\.07,p<10−12p<10^\{\-12\},W=0\.282W=0\.282\)\. The selected pair significantly outperformed most lower\-performing pairwise and single\-input models after Holm correction\. However, differences between the selected pair and either the full model or T1c\+T2w were not significant after correction\.

### 4\.2Post\-hoc Attribution

![Refer to caption](https://arxiv.org/html/2607.15396v1/figures/shapley_simple_boxplots_raw.png)Figure 3:Test\-set Shapley analysis of the four\-input model\. T2\-FLAIR and T1c had the largest input contributions, and their pairwise interaction was the strongest\.Shapley analysis ranked T2\-FLAIR first \(0\.336; 95% CI 0\.298–0\.374\) and T1c second \(0\.258; 95% CI 0\.203–0\.311\)\. Their interaction was also largest \(0\.307; 95% CI 0\.241–0\.373; Fig\.[3](https://arxiv.org/html/2607.15396#S4.F3)\)\. Pair interactions differed across the six combinations \(χ52=71\.22\\chi^\{2\}\_\{5\}=71\.22,p<10−13p<10^\{\-13\}, Kendall’sW=0\.509W=0\.509\)\. This independent analysis agreed with the PID\-selected input pair\.

## 5Discussion

### 5\.1PID Identified the Strongest Input Pair

The central result is that the MMI\-PID score selected the strongest two\-input configuration before segmentation training\. T1c\+T2\-FLAIR ranked first by the information score and was independently identified as the best pair in the 11\-model evaluation\. This agreement supports the use of the proposed score as a pre\-training input pair selector\.

The selected model used half the input channels and retained 98\.5% of the full input model mean Dice\. Paired tests did not provide clear evidence of differences in Dice, HD95, sensitivity, a larger study is required to reliably draw conclusions\. However, our work only serves as a proof of concept and is limited by only 4 input contrasts and small dataset splits\.

### 5\.2The Selected Pair Matches the Imaging Task

The selection of T1c and T2\-FLAIR is consistent with the reported regions\. T1c directly emphasizes enhancing tumor, while T2\-FLAIR is sensitive to edema and surrounding tissue abnormality\. Their combination therefore covers two distinct components of the segmentation target\. The region\-specific PID scores reflected this pattern as \(Table[1](https://arxiv.org/html/2607.15396#S4.T1)\)\. T1c\+T2\-FLAIR achieved the highest score for all three evaluated regions, including edema, enhancing tumor, and necrotic tumor\.

Shapley attribution provided a separate model\-based audit\. T2\-FLAIR and T1c had the two largest individual contributions, and their interaction was larger than every other pair interaction\. However, PID and Shapley address different questions\. PID ranks inputs before downstream training, whereas Shapley measures how a specific trained full model uses its inputs under a defined missing\-input baseline\. Their agreement strengthens the interpretation that T1c\+T2\-FLAIR captures both strong individual signal and useful cross\-input dependence\.

### 5\.3Implications for Resource\-Constrained Training

Selection occurs before costly segmentation training\. Fornncandidate inputs, the method trainsnnshallow autoencoders and evaluates\(n2\)\\binom\{n\}\{2\}pair scores, avoiding a separate downstream model for every pair\.

These pre\-training steps are comparatively lightweight and reusable across architectures\. Reducing the number of inputs also lowers data\-transfer, preprocessing, latency, augmentation, memory, and computation costs\.

## 6Conclusion

In this work, our framework using shallow autoencoder encoding, latent quantization, and a region\-aware MMI\-PID score identified the strongest two\-input configuration before segmentation training\. The framework selected T1c\+T2\-FLAIR, which halved the number of input channels, retained 98\.5% of the full\-model test Dice, and was independently supported by Shapley attribution\. This study provides a focused proof of concept for PID based input selection as a practical and interpretable pre\-training strategy for reducing multi\-contrast 3D input complexity\. Future work should evaluate its robustness in larger cohorts and settings with more than four candidate inputs\.

## References

- \[1\]\(2015\)Exploration of synergistic and redundant information sharing in static and dynamical gaussian systems\.Physical Review E91\(5\),pp\. 052802\.External Links:[Document](https://dx.doi.org/10.1103/PhysRevE.91.052802)Cited by:[§1](https://arxiv.org/html/2607.15396#S1.p4.2),[§2\.1](https://arxiv.org/html/2607.15396#S2.SS1.p3.5)\.
- \[2\]N\. Bertschinger, J\. Rauh, E\. Olbrich, J\. Jost, and N\. Ay\(2014\)Quantifying unique information\.Entropy16\(4\),pp\. 2161–2183\.External Links:[Document](https://dx.doi.org/10.3390/e16042161)Cited by:[§1](https://arxiv.org/html/2607.15396#S1.p4.2)\.
- \[3\]A\. S\. Chopra, C\. Neher, T\. Ren, J\. E\. Heras Rivera, H\. Jahanian, and M\. Kurt\(2026\)SFL\-Net: source\-factorized latent representation learning for multi\-contrast MRI to tau\-PET synthesis\.arXiv preprint arXiv:2602\.22545\.External Links:[Document](https://dx.doi.org/10.48550/arXiv.2602.22545),[Link](https://arxiv.org/abs/2602.22545)Cited by:[§1](https://arxiv.org/html/2607.15396#S1.p4.2)\.
- \[4\]Ö\. Çiçek, A\. Abdulkadir, S\. S\. Lienkamp, T\. Brox, and O\. Ronneberger\(2016\)3D U\-Net: learning dense volumetric segmentation from sparse annotation\.InMedical Image Computing and Computer\-Assisted Intervention – MICCAI 2016,Lecture Notes in Computer Science, Vol\.9901,pp\. 424–432\.External Links:[Document](https://dx.doi.org/10.1007/978-3-319-46723-8%5F49)Cited by:[§1](https://arxiv.org/html/2607.15396#S1.p1.1),[§2\.2](https://arxiv.org/html/2607.15396#S2.SS2.p1.3)\.
- \[5\]T\. M\. Cover and J\. A\. Thomas\(2006\)Elements of information theory\.2 edition,John Wiley & Sons,Hoboken, NJ\.Cited by:[§2\.1](https://arxiv.org/html/2607.15396#S2.SS1.p3.5)\.
- \[6\]J\. E\. Heras Rivera, D\. K\. Low, X\. Xiong, J\. J\. Ruzevick, D\. D\. Child, W\. Yim, M\. Kurt, and A\. Ben Abacha\(2026\)CoRe\-BT: a multimodal radiology\-pathology\-text benchmark for robust brain tumor typing\.arXiv preprint arXiv:2603\.03618\.External Links:[Document](https://dx.doi.org/10.48550/arXiv.2603.03618),[Link](https://arxiv.org/abs/2603.03618)Cited by:[§1](https://arxiv.org/html/2607.15396#S1.p1.1),[§1](https://arxiv.org/html/2607.15396#S1.p4.2),[§3](https://arxiv.org/html/2607.15396#S3.p1.1)\.
- \[7\]G\. E\. Hinton and R\. R\. Salakhutdinov\(2006\)Reducing the dimensionality of data with neural networks\.Science313\(5786\),pp\. 504–507\.External Links:[Document](https://dx.doi.org/10.1126/science.1127647)Cited by:[§2\.1](https://arxiv.org/html/2607.15396#S2.SS1.p2.7)\.
- \[8\]S\. Holm\(1979\)A simple sequentially rejective multiple test procedure\.Scandinavian Journal of Statistics6\(2\),pp\. 65–70\.Cited by:[§3](https://arxiv.org/html/2607.15396#S3.p3.1)\.
- \[9\]F\. Isensee, P\. F\. Jaeger, S\. A\. A\. Kohl, J\. Petersen, and K\. H\. Maier\-Hein\(2021\)nnU\-Net: a self\-configuring method for deep learning\-based biomedical image segmentation\.Nature Methods18,pp\. 203–211\.External Links:[Document](https://dx.doi.org/10.1038/s41592-020-01008-z)Cited by:[§1](https://arxiv.org/html/2607.15396#S1.p1.1)\.
- \[10\]S\. Lloyd\(1982\)Least squares quantization in PCM\.IEEE Transactions on Information Theory28\(2\),pp\. 129–137\.External Links:[Document](https://dx.doi.org/10.1109/TIT.1982.1056489)Cited by:[§2\.1](https://arxiv.org/html/2607.15396#S2.SS1.p2.7)\.
- \[11\]I\. Loshchilov and F\. Hutter\(2017\)SGDR: stochastic gradient descent with warm restarts\.InInternational Conference on Learning Representations,Cited by:[§2\.2](https://arxiv.org/html/2607.15396#S2.SS2.p2.4)\.
- \[12\]I\. Loshchilov and F\. Hutter\(2019\)Decoupled weight decay regularization\.InInternational Conference on Learning Representations,Cited by:[§2\.2](https://arxiv.org/html/2607.15396#S2.SS2.p2.4)\.
- \[13\]B\. H\. Menzeet al\.\(2015\)The multimodal brain tumor image segmentation benchmark \(BRATS\)\.IEEE Transactions on Medical Imaging34\(10\),pp\. 1993–2024\.External Links:[Document](https://dx.doi.org/10.1109/TMI.2014.2377694)Cited by:[§1](https://arxiv.org/html/2607.15396#S1.p1.1)\.
- \[14\]H\. Peng, F\. Long, and C\. Ding\(2005\)Feature selection based on mutual information: criteria of max\-dependency, max\-relevance, and min\-redundancy\.IEEE Transactions on Pattern Analysis and Machine Intelligence27\(8\),pp\. 1226–1238\.External Links:[Document](https://dx.doi.org/10.1109/TPAMI.2005.159)Cited by:[§1](https://arxiv.org/html/2607.15396#S1.p3.1)\.
- \[15\]T\. Ren, J\. E\. Heras Rivera, H\. Oswal, Y\. Pan, A\. S\. Chopra, J\. J\. Ruzevick, and M\. Kurt\(2025\)Here comes the explanation: a shapley perspective on multi\-contrast medical image segmentation\.InJoint Proceedings of the xAI 2025 Late\-Breaking Work, Demos and Doctoral Consortium,CEUR Workshop Proceedings, Vol\.4017,pp\. 185–192\.External Links:[Link](https://ceur-ws.org/Vol-4017/paper_24.pdf)Cited by:[§1](https://arxiv.org/html/2607.15396#S1.p4.2),[§3](https://arxiv.org/html/2607.15396#S3.p3.1)\.
- \[16\]Y\. Sadegheih and D\. Merhof\(2024\)Segmentation of brain metastases in MRI: a two\-stage deep learning approach with modality impact study\.InPredictive Intelligence in Medicine,I\. Rekik, E\. Adeli, S\. H\. Park, and C\. Cintas \(Eds\.\),Lecture Notes in Computer Science, Vol\.15155,pp\. 196–206\.External Links:[Document](https://dx.doi.org/10.1007/978-3-031-74561-4%5F17)Cited by:[§1](https://arxiv.org/html/2607.15396#S1.p4.2)\.
- \[17\]L\. S\. Shapley\(1953\)A value fornn\-person games\.InContributions to the Theory of Games II,H\. W\. Kuhn and A\. W\. Tucker \(Eds\.\),Annals of Mathematics Studies,pp\. 307–317\.Cited by:[§3](https://arxiv.org/html/2607.15396#S3.p3.1)\.
- \[18\]C\. Westphal, S\. Hailes, and M\. Musolesi\(2025\)Partial information decomposition for data interpretability and feature selection\.InProceedings of the 28th International Conference on Artificial Intelligence and Statistics,Y\. Li, S\. Mandt, S\. Agrawal, and E\. Khan \(Eds\.\),Proceedings of Machine Learning Research, Vol\.258,pp\. 1873–1881\.External Links:[Link](https://proceedings.mlr.press/v258/westphal25a.html)Cited by:[§1](https://arxiv.org/html/2607.15396#S1.p4.2),[§2\.1](https://arxiv.org/html/2607.15396#S2.SS1.p4.3)\.
- \[19\]P\. L\. Williams and R\. D\. Beer\(2010\)Nonnegative decomposition of multivariate information\.arXiv preprint arXiv:1004\.2515\.External Links:[Document](https://dx.doi.org/10.48550/arXiv.1004.2515),[Link](https://arxiv.org/abs/1004.2515)Cited by:[§1](https://arxiv.org/html/2607.15396#S1.p4.2),[§2\.1](https://arxiv.org/html/2607.15396#S2.SS1.p3.5)\.
- \[20\]P\. Wollstadt, S\. Schmitt, and M\. Wibral\(2023\)A rigorous information\-theoretic definition of redundancy and relevancy in feature selection based on \(partial\) information decomposition\.Journal of Machine Learning Research24\(131\),pp\. 1–44\.External Links:[Link](https://www.jmlr.org/papers/v24/21-0482.html)Cited by:[§1](https://arxiv.org/html/2607.15396#S1.p4.2),[§2\.1](https://arxiv.org/html/2607.15396#S2.SS1.p4.3)\.
- \[21\]T\. Zhou, S\. Canu, P\. Vera, and S\. Ruan\(2021\)Feature\-enhanced generation and multi\-modality fusion based deep neural network for brain tumor segmentation with missing MR modalities\.Neurocomputing466,pp\. 102–112\.External Links:[Document](https://dx.doi.org/10.1016/j.neucom.2021.09.032)Cited by:[§1](https://arxiv.org/html/2607.15396#S1.p4.2)\.