The system rests on three components. First, a DenseNet-121 classifier is trained on the BreaKHis dataset of 7,909 breast histology images using knowledge distillation, in which a compact student network learns from a larger InceptionV3 teacher. On a held-out test set split at the patient level to prevent data leakage, the student achieved an AUC of 0.9989 and 99.2 percent accuracy, actually surpassing its teacher. Second, the GRALIS algorithm, whose name stands for Gradient-Riesz Averaged Locally-Integrated Shapley, divides each image into roughly 25 spatially coherent tissue segments called superpixels and assigns each one an importance score indicating whether it raised or lowered the malignancy probability. Third, a deterministic rule-based engine converts these scores into a clinical report organized in the familiar Subjective, Objective, Assessment, and Plan format used throughout medicine.

What distinguishes GRALIS from better-known explanation methods is its technical positioning within the explainable AI design space. GradCAM, the most widely used technique in medical imaging, produces low-resolution heat maps that must be upsampled and can highlight regions the model never actually consulted. Integrated Gradients guarantees mathematical completeness, meaning its per-pixel scores sum exactly to the model’s output change, but its dense pixel maps bear little relation to anatomically meaningful structures. KernelSHAP and LIME operate on superpixels but rely on post-hoc convergence detection, so users cannot know in advance how many samples an approximation requires. GRALIS, by contrast, offers an a priori Monte Carlo sample-size bound: the number of perturbation samples needed for a target error can be computed before the attribution run, after a brief calibration pass.

The algorithmic novelty lies in what the author calls coalition-conditioned path integration. Rather than evaluating gradients along a single global path as standard Integrated Gradients does, GRALIS computes gradients along paths specific to each coalition of segments, capturing how the model’s curvature changes depending on which tissue regions are present or absent. This comes at a deliberate cost: the locality kernel that makes region-coherent attribution possible is formally incompatible with exact completeness, so GRALIS accepts a non-zero empirical residual of 0.40 in exchange for maps that pathologists can actually inspect. The companion theoretical work proves this incompatibility is structural, not incidental, meaning no method can simultaneously provide perfect efficiency and LIME-style locality.

Benchmark results on the full BreaKHis test set of 1,187 images reveal a nuanced trade-off profile. On independent pixel-level deletion faithfulness, GRALIS ranked second of six methods with a deletion AUC of 0.704, trailing only Integrated Gradients at 0.678 and statistically outperforming GradCAM, GradCAM++, KernelSHAP, and LIME. On the ROAD metric, it again ranked second while achieving the largest discrimination gap between relevant and irrelevant regions among the three evaluated methods. When the comparison was made granularity-fair, masking entire superpixels rather than individual pixels, Integrated Gradients’ advantage shrank dramatically from 0.119 to 0.019, suggesting that much of its pixel-level superiority reflects dense attribution granularity rather than genuinely better region identification.

Cross-dataset evaluation sharpened the picture in an unusually honest way. On the IDC breast cancer dataset of 499 small patches, using the same frozen backbone with a newly trained linear classification head, GRALIS ranked first on both Deletion AUC and ROAD MoRF, beating Integrated Gradients by roughly 17 percent on deletion faithfulness. But on PatchCamelyon, a set of 96-by-96-pixel lymph node sections, GRALIS dropped to last place. The author attributes this to a granularity mismatch: with the segment count fixed at 30, each superpixel covers approximately 368 native pixels, far coarser than the 5-to-15-pixel cellular features that distinguish tumor in lymph node tissue. The explanation remains a hypothesis pending a dedicated ablation, but the contrasting outcomes illustrate that superpixel-based attribution succeeds where region scale matches the discriminative anatomy, not universally.

The reporting engine itself proved remarkably reliable. Applied to all 1,187 test images in batch mode, it produced 1,187 syntactically complete SOAP reports with zero execution failures and no external dependencies, operating fully offline in under a hundredth of a second per report. Of these, 1,175, or 98.99 percent, corresponded to correct classifier predictions. The twelve discordant cases, six false positives and six false negatives, were automatically flagged with priority review alerts embedded in the Assessment and Plan fields, and were identified retrospectively using test labels. In one illustrative malignant case at 400-times magnification, the report named the top three tissue regions driving the prediction, each occupying roughly three percent of the image, alongside confidence and attribution quality metrics.

Perhaps the most thought-provoking result is a small expert perception study with four anatomopathologists reviewing 60 cases. Preference for GRALIS maps over GradCAM was strong in aggregate, 67 percent versus 25 percent, yet individual raters diverged wildly: one pathologist rated the clinical utility of the reports at a mean of 4.73 out of 5 while another scored the identical material at 1.47. Free-annotation tasks showed concordance ranging from 12 percent on benign cases to 83 percent on missed malignancies, with disagreement driven largely by annotation style. The author frames this as a methodological warning rather than a verdict on the system: with such variability, perceived explanation utility may simply not constitute a stable ground truth in histopathology, complicating the design of future human evaluations of explainable AI.

The study is notable for its unusual transparency about its own limits. The author declares that the supplementary ExpiScore metric shares authorship with the work and should be weighted accordingly, relying instead on independent deletion and ROAD benchmarks for primary claims. An initial protocol error in the PatchCamelyon validation was detected, corrected, and reported with the finding that all faithfulness rankings were unchanged. No clinical efficacy claims are made anywhere in the paper; the framework is presented as an exploratory engineering prototype, not a diagnostic device. That candor, combined with the strict separation between visual evidence and generated language, positions GRALIS-Report less as a finished clinical tool than as a template for how auditable, reproducible AI reporting in pathology might be built, where every number in a report can be traced back to a stored attribution record and every design trade-off is stated rather than hidden.

Subject of Research: Explainable AI and auditable report generation for breast cancer histopathology images

Article Title: GRALIS-Report: auditable region-level attribution and structured clinical report generation for breast cancer histology

Article References: Fanale, R. (2026). GRALIS-Report: auditable region-level attribution and structured clinical report generation for breast cancer histology. Frontiers in Imaging, 5, Article 1885292. https://doi.org/10.3389/fimag.2026.1885292

Image Credits: AI Generated

DOI: 10.3389/fimag.2026.1885292

Keywords: explainable AI, breast cancer histology, attribution maps, Shapley values, clinical report generation, BreaKHis, knowledge distillation, superpixels, digital pathology, faithfulness benchmarking, deletion AUC, PATCHCamelyon