Source-linked AI summary
MUFASA: An Information Utility-Aware Preprocessing Framework for Reliable Model Reasoning in Computational Pathology
Rathinaraja Jeyaraj, Barathi Subramanian, Songmi Noh, Mitchell N. Peterson, Terry Guo, George A. Fisher, Nigam H. Shah, Curtis P. Langlotz, Thomas J. Montine, Jeanne Shen
TL;DR
Existing WSI preprocessing can retain artifacts and low-utility regions or remove relevant tissue, limiting reliable model reasoning. MUFASA combines progressive artifact suppression with utility-aware tile selection for H&E-stained WSI. Across diagnostic, biomarker, and survival tasks, it consistently improved downstream performance while reducing artifact-associated attribution.
Problem
Existing pipelines may retain artifacts and low-utility regions or discard diagnostically relevant tissue, affecting downstream model performance, reliability, and robustness.
Method
MUFASA progressively masks artifacts, filters tiles, stratifies utility with reconstruction loss, and selectively recovers biologically relevant tissue.
Results
MUFASA consistently improved downstream performance across diagnostic, biomarker-status, and survival tasks while reducing artifact-driven model attribution.
Takeaways & Limitations
WSI preprocessing critically affects model reliability and interpretability, so accurate slide-level predictions should be assessed alongside the anatomical plausibility of model attention.
Takeaways & Limitations
Evaluation focused primarily on H&E-stained WSIs from three sources, while additional dataset-specific artifacts, scanner variations, institutional differences, and non-H&E stains were not systematically assessed.
Abstract
from arXiv · showhide
Reliable computational pathology depends on preprocessing methods that identify informative tissue regions while excluding artifacts and low-utility regions from whole-slide images (WSI). However, existing preprocessing pipelines often retain such regions or discard diagnostically relevant tissue, thereby limiting downstream model performance, reliability, and robustness across heterogeneous cohorts. Here, we systematically evaluate how these regions affect downstream AI model performance across multiple clinically relevant applications and introduce MUFASA, a generalizable, information utility-aware preprocessing framework for H&E-stained WSI that excludes artifacts and low-utility regions while preserving biologically meaningful tissue. MUFASA integrates slide-level artifact masking, stain-aware tile filtering, reconstruction-based utility stratification of tiles, and targeted recovery of tissue tiles that are over-filtered by earlier phases. Across tumor diagnosis, tumor subtyping, biomarker status prediction, and survival prognostication tasks in diverse cancer cohorts, MUFASA consistently improves downstream model performance relative to widely used preprocessing baselines. These gains are accompanied by reduced artifact-associated attribution in model heatmaps, indicating improved alignment between retained tissue and model attention. Our findings establish WSI preprocessing as a critical determinant of downstream model performance and validity, revealing that even accurate predictions can conceal important failure modes stemming from anatomically implausible reasoning driven by retained artifact-containing and low information-utility tiles.
1. Introduction
Computational pathology pipelines can retain artifacts and low-utility regions or discard diagnostically relevant tissue, making WSI preprocessing a challenge for reliable model reasoning. MUFASA addresses this gap with progressive, information utility-aware filtering that suppresses artifacts while preserving biologically meaningful tissue.
- WSI preprocessing commonly tiles slides, extracts tile features, and aggregates them into slide- or patient-level predictions with multiple instance learning models.
- Retained artifacts can encourage reliance on spurious visual cues and shift model attribution away from biologically meaningful morphology, even when predictions are correct.
- Heuristic and widely used preprocessing pipelines can misclassify high-contrast artifacts, retain fine-grained artifacts, or discard diagnostically relevant tissue.
- Binary tissue-versus-artifact classification overlooks the continuous, task-dependent nature of tile utility and can generalize poorly under domain shift.
- MUFASA progressively removes background and artifacts, stratifies retained tiles by reconstruction-based utility, and recovers visually subtle but biologically relevant compartments.
- The study evaluates preprocessing effects across tumor diagnosis and subtyping, biomarker prediction, and survival prognostication using diverse cancer cohorts.
2.1. Tasks and datasets
The study tests preprocessing robustness across four clinically relevant WSI-level tasks using cohorts that span multiple cancer types, institutions, scanners, staining profiles, artifact types, and image resolutions.
- Experiments covered tumor diagnosis and subtyping, molecular biomarker status prediction, and survival prognostication across heterogeneous WSI cohorts.
- Diagnostic classification included CAMELYON16 metastasis detection and TCGA lung adenocarcinoma versus squamous cell carcinoma subtyping.
- Prognostic modeling used TCGA-STAD with 364 gastric cancer cases and TCGA-LUAD with 376 cases.
- The datasets collectively represent variation in organs, institutions, scanners, H&E staining profiles, artifact types, and image resolutions.
2.2. Baseline comparison, tiling and feature extraction
MUFASA was compared with CLAM, Trident, and Histolab using generated WSI tissue masks and standardized downstream tile extraction, with Set1 tiles used for MUFASA analyses.
- MUFASA was compared against CLAM, Trident, and Histolab, with each method generating WSI-level tissue masks for downstream tile extraction.
- MUFASA used Set1 high-utility tiles for downstream tasks, while Set2 and Set3 were excluded because of limited task-relevant morphologic content.
- Tiles were extracted using resolution-normalized settings to maintain consistent physical tissue coverage across slides.
2.3. MIL benchmarking for classification and biomarker status prediction
The study benchmarks preprocessing across diverse MIL architectures and evaluates whether MUFASA’s gains are consistent across models and metrics rather than driven by isolated results.
- Thirteen MIL models spanning pooling, attention, transformers, dual streams, graphs, hard-instance mining, and state-space modeling were evaluated for classification tasks.
- The benchmark included ACC, BACC, Macro-AUC, and Macro-F1 to assess performance across preprocessing methods.
- Paired Wilcoxon signed-rank tests compared MUFASA with CLAM, Histolab, and Trident across models and evaluation metrics.
2.4. Impact on WSI classification
Across tumor diagnosis and subtyping, MUFASA improved performance across diverse MIL models while retaining diagnostically relevant tissue and reducing artifact-driven attention. These gains were observed across classification metrics, latent feature organization, and attention localization.
- Performance: MUFASA consistently outperformed CLAM, Histolab, and Trident across 13 MIL models for CAMELYON16 tumor classification and NSCLC tumor subtyping.The gains were distributed across model families, indicating that preprocessing rather than architecture-specific optimization primarily accounted for the improvement.
- Performance: +2.34 ACC points was MUFASA’s improvement over CLAM on CAMELYON16, with average ACC reaching 78.16 versus 75.82.MUFASA also improved ACC over Histolab and Trident by +4.12 and +4.61 points, respectively.
- Performance: +0.061 Macro-AUC was MUFASA’s mean improvement over Trident on CAMELYON16, compared with +0.035 over CLAM and +0.057 over Histolab.On NSCLC, the corresponding mean gains were +0.024, +0.017, and +0.021, respectively.
- Representation: MUFASA produced cleaner class separation than CLAM or Histolab in t-SNE feature spaces, whereas Trident’s sharper separation coincided with aggressive removal of diagnostically relevant regions.The qualitative feature-space analysis was consistent with artifact-driven variance in the permissive baselines and performance degradation under aggressive filtering.
- Interpretability: MUFASA concentrated attention on diagnostically relevant tissue while minimizing attention to uninformative regions across representative CAMELYON16 and TCGA-LUSC slides.On TCGA-LUSC, MUFASA improved APS by 0.21 over CLAM, 0.215 over Histolab, and 0.207 over Trident while retaining 96.23% of relevant tiles and removing 100% of uninformative tiles.
2.5. Impact on Survival Analysis
MUFASA improved prognostic modeling across gastric and lung adenocarcinoma cohorts using 13 MIL baselines. Its preprocessing was also associated with more consistent survival-risk separation and reduced attention to artifact and low-utility regions.
- Survival prediction: MUFASA consistently outperformed CLAM, Histolab, and Trident across all evaluated MIL models for survival prediction.The evaluation covered TCGA-LUAD and TCGA-STAD using fivefold cross-validation and held-out concordance index.
- Survival prediction: +0.016 to +0.039% was MUFASA’s C-index improvement over CLAM across models on TCGA-LUAD.On TCGA-STAD, the corresponding gains ranged from +0.014 to +0.038%.
- Risk stratification: MUFASA yielded the strongest and most consistent separation of survival trajectories in both TCGA-STAD and TCGA-LUAD under median-risk stratification.Kaplan–Meier curves were generated with the best-performing survival model, OTSurv, as an orthogonal assessment.
- Attention localization: Attention heatmaps on artifact-containing TCGA-STAD slides showed that preprocessing changed which regions entered the prognostic model and received attention.The analysis included pen marks, tissue folds, and blurry regions, with low-utility tiles largely removed before MIL processing under MUFASA.
- Preprocessing reliability: The stress test quantified a balanced preprocessing score of 94% for MUFASA, compared with 70% for Trident and 38–39% for CLAM and Histolab.CLAM and Histolab favored tissue retention with limited artifact removal, whereas Trident favored artifact removal with reduced tissue retention.
2.6. Pathologist evaluation and quality control via regional stress-test on 2048 × 2048 ROIs
The regional stress-test compared preprocessing masks on pathologist-selected challenge ROIs across six TCGA cohorts. MUFASA best balanced artifact suppression with preservation of tissue-rich and fragmented regions.
- Regional stress-test: The stress-test evaluated artifact-free tissue, pen marks, blur, contaminants, and fragmented or low-tissue regions across 72 ROIs.ROIs were manually selected by a pathologist from six TCGA cohorts and partitioned into an 8×8 grid.
- Regional stress-test: Identical ROI coordinates enabled direct comparison of CLAM, Histolab, Trident, and MUFASA independently of ROI selection.
- Qualitative findings: CLAM and Histolab frequently retained pen-marked, blurred, and contaminant-dense regions, while permissive pipelines retained 93.9–100.0% of annotated uninformative tiles in the NSCLC cohort.
- Quantitative comparison: MUFASA achieved 93–99% artifact removal, 88–92% tissue preservation, and 92–95% balanced performance across cohorts.Its aggregate preprocessing reliability was 94%, compared with 70% for Trident and 38–39% for CLAM and Histolab.
3. Discussion
The discussion frames preprocessing as a determinant of computational pathology reliability because retained artifacts and low-utility tiles can distort MIL reasoning. MUFASA addresses this by balancing artifact suppression with information preservation, while several scope limitations remain.
- Implications: WSI preprocessing affected downstream reliability, interpretability, and efficiency across diagnostic, classification, biomarker, and survival tasks.
- Implications: Correct slide-level predictions can conceal anatomically implausible attention, so predictive accuracy alone is insufficient for assessing preprocessing quality.
- Operating point: MUFASA balanced strong artifact suppression with tissue coverage, accompanying improved Macro AUC, Macro F1, C-index, latent separation, attention maps, and ROI stress-test performance.
- Conceptual contribution: Unlike fixed-rule and segmentation-focused alternatives, MUFASA treats preprocessing as progressive utility-aware filtering with modular extensibility across heterogeneous cohorts.
- Limitations: Evaluation focused primarily on H&E-stained WSIs from TCGA, CAMELYON16, and Stanford, leaving additional artifacts, scanners, institutions, and non-H&E stains insufficiently tested.
- Limitations: The Set1-only strategy and single-reader analyses on small slide subsets limit conclusions about broader task applicability and annotation reliability.
- Conclusion: The study concludes that information utility-aware preprocessing can improve downstream performance while reducing artifact-driven attribution and strengthening anatomic validity.
4. Methods
MUFASA uses four modular phases to create utility-aware tissue masks from H&E-stained WSIs. The pipeline combines artifact suppression, stain-aware filtering, reconstruction-based stratification, and recovery of tissue vulnerable to over-filtering.
- Framework: MUFASA progressively removes background and artifacts while preserving biologically meaningful tissue through four modular preprocessing phases.Targets include air bubbles, dust, debris, pen marks, folds, blur, and scanner- or stain-related artifacts.
- Phase 1: Pen marks are non-biological ink annotations that enter WSI tile extraction as high-contrast structures and can occur in multiple colors across institutions.
- Phase 2: Phase 2 transforms RGB intensities into optical-density space for stain-aware tile filtering that is less sensitive to illumination and color variation.
- Phase 2: Tiles were discarded when mean eosin-channel intensity was at most 0.01, dark regions exceeded 25%, edge score fell below 3%, uniform regions exceeded 70%, or background exceeded 85%.
- Phase 3: Phase 3 uses an autoencoder trained on uninformative tiles so artifact-like patterns have low reconstruction loss while biologically meaningful tissue has comparatively higher loss.
- Phase 3: Inference stratifies tiles into Sets 1–4 using reconstruction loss and uniform-region scores, with Set1 representing tissue-dense regions considered relevant for downstream analysis.
- Threshold selection: The selected loss thresholds were t1=0.002, t2=0.0015, and t3=0.001, chosen by data-driven ablation and Pareto-efficient artifact-preservation trade-offs.
- Phase 4: Phase 4 addresses over-filtering of adipose- and mucin-rich tissue, which can resemble background or artifacts because of weak stain contrast and optically clear areas.
Supplementary Materials
Across tumor diagnosis, tumor subtyping, biomarker-status prediction, and survival prognostication, MUFASA generally outperformed alternative preprocessing pipelines while keeping feature extraction and MIL training settings fixed. The comparisons evaluate whether artifact-aware tissue selection improves downstream performance across architectures and cohorts.
- Tumor subtyping: MUFASA produced the best overall performance across architectures for NSCLC tumor subtyping.All MIL models used identical feature extraction and optimization settings; only the preprocessing pipeline differed.
- Tumor diagnosis: MUFASA yielded the strongest and most consistent improvements for tumor diagnosis on CAMELYON16 across preprocessing comparisons.Accuracy, balanced accuracy, Macro-AUC, and Macro-F1 were reported across five random runs with downstream settings held constant.
- Biomarker-status prediction: MUFASA consistently improved MSS/MSI prediction performance relative to baseline preprocessing pipelines across independent colorectal cancer cohorts.The evaluation used accuracy, balanced accuracy, Macro-AUC, and Macro-F1, with preprocessing as the only varying component.
- Survival prognostication: MUFASA consistently improved prognostic performance relative to CLAM-style preprocessing, Histolab, and Trident across diverse MIL model families.Survival results were reported as held-out C-index means and standard deviations on TCGA-LUAD and TCGA-STAD.
Model TCGA-LUAD TCGA-STAD
The supplementary results include detailed performance values for MUFASA-based and alternative preprocessing pipelines, alongside survival evaluations and phase-wise or utility-aware visual audits. These materials show how preprocessing outputs and retained tile sets are organized for downstream comparison.
- Biomarker-status prediction: The Stanford colorectal MSS/MSI table reports accuracy values for multiple MIL architectures under CLAM, Histolab, Trident, and MUFASA preprocessing.The listed rows include Max-pooling, Mean-pooling, ABMIL, CLAM-MB, and TransMIL, with mean±s.d. values.
- Preprocessing audit: The phase-wise audit trail documents preprocessing inputs and outputs, rejected regions, and representative artifacts such as pen marks, blur, dust, and staining irregularities.It identifies slide-level artifact removal as an early MUFASA phase.
- Utility-aware masks: Utility-aware tissue masks remap Set1, Set2, and Set3 tiles to slide coordinates and aggregate them into a combined tissue mask.The overlays depict suppression of artifact-dominant and low-utility regions while preserving retained tissue.