PaperlayerView exact revision on arXiv

Source-linked AI summary

Evidence Field for Histopathology: Cross-Centre Evaluation and Replayable Spatial Decisions

Isarar Siddique, Raiyan Siddique, Arshi Niyaz, Jamyang Ugyen Tshomo, Atish Raza

arXiv:2610.21115v1eess.IV

TL;DR

Histopathology classifiers need to transfer across acquisition sites while exposing decision paths that can be checked independently. Evidence Field couples learned image structure with spatial covering programs and replayable predictions. It leads the implemented CNN and ViT comparators on both held-out centres, while the comparison and evidence audit have defined scope.

Abstract

from arXiv · show

Histopathology classifiers must transfer across acquisition sites while exposing evidence that can be checked independently. Evidence Field couples learned image measurements and spatial classification with a native covering-program record. We evaluate this architecture on all 455,954 released CAMELYON17-WILDS patches, using patient-separated source partitions and two held-out hospital centres. Eighteen from-scratch fits compare four Evidence Field objectives with residual convolutional and compact vision-transformer classifiers. Checkpoints, temperatures and review thresholds are selected using source validation. The source-selected paired-cross-entropy ensemble achieves areas under the receiver operating characteristic curve of 0.9356 and 0.9453, with accuracies of 84.46% and 86.77% at the two held-out centres. Corresponding convolutional values are 0.7352 and 0.7656, and transformer values are 0.8502 and 0.7924. Evidence Field has 10,550 trainable parameters per member and additionally returns selected covers, signed region and interaction terms, contextual dependencies and independently reconstructible predictions. Patient-cluster intervals, calibration, source-fixed review policies and recorded training resources characterize the results. A 340-image, three-member audit reconstructs probability mixtures and 6,120 class programs under three rigid views. The evaluation supports cross-centre discrimination and reproducible computation for these implemented models; unequal family tuning, previously examined public targets and limited patient groups constrain broader interpretation.

I. INTRODUCTION

The paper tests whether histopathology classifiers can transfer across centres while exposing independently checkable decision paths. Evidence Field couples learned measurements with spatial covering programs; the study situates this approach among distinct target, scale, and evidence settings.

I. INTRODUCTION

  • Evidence Field makes classification computation explicit by using a legal spatial-covering program to select the signed contributions that generate logits.The study compares four EF objectives with CNN and ViT classifiers while holding source images, epochs, and seeds fixed.Source: §I · p. 1

A. Nearest mechanisms

  • EF couples referenced signatures and cross-scale scores to selected covers and independently reconstructible optima.The application evaluates cross-centre classification, calibration, cost, and native evidence; the accompanying technical manuscript defines the core.Source: A. Nearest mechanisms · p. 1

A. Targets and scale

  • Tumour detection, molecular prediction, and outcomes are distinct endpoints, while foundation models use pretraining scales and supervision outside this from-scratch comparison.Native 96×96 lymph-node patches also differ from rescaled benchmarks.Source: A. Targets and scale · p. 1

B. Acquisition and evidence

  • EF’s supports are computational regions, distinct from biological localization; reconstruction and hospital-shift auditing are also distinct from clinical screening and referral validation.Source: B. Acquisition and evidence · p. 1

III. DATA SOURCE, LABEL DEFINITION AND QUALITY

The study uses released CAMELYON17-WILDS patches with source patients separated from held-out hospital-centre tests. Its central-window tumour label and data-quality checks define the evaluation’s reference and scope.

A. Origin and access

  • CAMELYON17-WILDS releases 455,954 native 96×96 RGB patches; centres 0/3/4 define the source and centres 1/2 supply evaluation only.Source train, validation, and internal-test patients are disjoint; the centre identifiers have no authenticated institution-name mapping.Source: A. Origin and access · p. 1

B. Reference label

  • A patch is positive when its central 32×32 window contains tumour, although a negative patch can contain tumour outside that window.The reference annotations were revised using cytokeratin immunohistochemistry; isolated-tumour-cell annotation was nonexhaustive.Source: B. Reference label · p. 2

C. Quality controls

  • All patches were retained after checks of shape, type, labels, assignments, coordinates, and decoded digests, with no patient overlap recorded.Exact-byte separation does not establish independence of overlapping neighbourhoods; scanner calibration, spacing, and expert focus ratings are unavailable.Source: C. Quality controls · p. 2

A. Optical observation contract

Evidence Field defines optical measurements from normalized RGB input and combines source-relative and current-image-relative response normalization. The current image supplies a reference without updating learned weights or using its label.

A. Optical observation contract

  • EF retains three absorption channels from RGB input rather than collapsing the image to one intensity.The input is 96×96 RGB scaled by exact uint8 division by 255; derivatives use pixel units because authenticated spacing is unavailable.Source: A. Optical observation contract · p. 2

B. Learned responses and same-image reference

  • Eight learned radial units mix source-standardized and current-image-standardized responses, with incoming-image statistics computed without its label or weight updates.Source normalization starts from 32 uniformly selected training patches; the reference is not presumed healthy, and robustness is assessed by centre transfer.Source: B. Learned responses and same-image… · p. 2

C. Within-region structural signatures

Evidence Field summarizes responses within regions using moments and finite characteristic-function signatures, then uses typed node and cross-scale scores to form signed contributions. Its descriptors encode structure, but the finite signature bank cannot distinguish every arrangement.

C. Within-region structural signatures

  • Each region accumulates response mass, centroid, covariance, and a finite learned characteristic-function bank; concatenated own, parent-difference, and root-difference descriptors yield D = 435.Arrangement-sensitive signatures augment low-order moments, but a finite bank cannot distinguish every possible arrangement.Source: C. Within-region structural signatures · p. 3

D. Typed node and cross-scale scores

  • Typed node and legal child-pair scores add four signed correction components, including child interaction and parent modulation.L is affine in mass-weighted child mean and absolute spread, while P is an affine parent-context term; these components describe computation, not causal effects.Source: D. Typed node and cross-scale scores · p. 3

E. Adaptive covering recursion

The depth-two registry provides two refinements per region, and dynamic programming selects the maximum-scoring legal covering program. The selected program covers the patch once and records its construction and terms.

E. Adaptive covering recursion

  • Two alternative refinements per region form the depth-two spatial registry, with each production partitioning its parent.Pair packing can leave some children unmatched; dynamic programming evaluates the alternatives.Source: E. Adaptive covering recursion · pp. 3–4
  • The exact maximum is over the finite legal registry, and every selected root program covers the patch once.A production combines disjoint child supports; the record includes region bounds, production and packing, selected type, anchors, correction terms, and bias.Source: E. Adaptive covering recursion · p. 4

F. Evidence identity

Evidence is displayed by assigning each unit’s score uniformly over its support, while preserving classifier arithmetic. Source partitions and selection are fixed before training, and public target histories make the evaluation an exploratory replication.

F. Evidence identity

  • The signed display conserves the class-logit difference, although global references let pixels outside a unit affect its score.The display assigns each unit’s score uniformly over its support.Source: F. Evidence identity · p. 4

A. Source partition and selection

  • Source training, validation, and internal testing use patient-separated partitions; centres 1 and 2 are held out for external testing.Centre 1 is evaluation-only here, so the study is not an exact official leaderboard reproduction; prior project use of the public targets makes it an exploratory replication.Source: A. Source partition and selection · p. 4

B. Models and optimization

The study compares from-scratch convolutional and compact transformer classifiers with four Evidence Field objectives across 18 fits. Source validation selects checkpoints, temperatures, and review cutoffs, while patient-cluster resampling characterizes external results.

B. Models and optimization

  • The from-scratch comparators are a ResNet18 and a compact ViT with four pre-normalized attention blocks and width 128.The ViT uses 144 non-overlapping 8×8 RGB patches, a class token, learned positions, eight attention heads, width-256 feed-forward units, and dropout 0.1.Source: B. Models and optimization · p. 4

C. Full-source fits

  • The six recipes each receive 16 epochs and three seeds, yielding 18 complete fits.Whole-patch quarter-turns and reflections are sampled; paired Evidence Field also receives an independently sampled geometric transform, treated as a label-preservation stress assumption rather than a guarantee.Source: C. Full-source fits · p. 4

D. Objectives and temperature

  • Four Evidence Field objectives share one architecture; paired cross-entropy is the headline recipe selected by mean three-seed source AUROC.Probability-regularized EF adds 0.05 symmetric KL, while signed-evidence EF further adds 0.02 normalized aligned all-class spatial-evidence distance; four EF recipes versus one per comparator is an unequal search budget.Source: D. Objectives and temperature · p. 4

E. Prediction, uncertainty and review

  • Headline probabilities average three raw seed probabilities, while calibrated ensembles average probabilities after source-only temperature scaling.Review cutoffs are source-validation quantiles; external coverage and retained error are reported without choosing external cutoffs, and percentile intervals use 2,000 patient-cluster draws with eight and nine external patient groups.Source: E. Prediction, uncertainty and review · p. 4

A. Paired patient uncertainty

The section reports exploratory patient-resampling intervals for external AUROC differences and describes what the comparator and independent-evidence results support. It also records boundaries on interpreting the comparison and training-cost measurements.

A. Paired patient uncertainty

  • Centre 1 EF–CNN and EF–ViT AUROC differences were [0.097, 0.444] and [0.005, 0.252]; Centre 2 differences were [0.108, 0.432] and [0.114, 0.299].These exploratory percentile intervals use 8 and 9 resampling units, with 2,000 AUROC draws defined per external comparison; they are not multiplicity-adjusted superiority tests.Source: A. Paired patient uncertainty · p. 4
  • Recorded full-source cost reports mean/range hours and maximum allocated device memory, but excludes uncheckpointed failed-attempt overhead and does not include all host/storage memory.Energy and isolated inference time were not measured; a small parameter count does not imply faster training.Source: A. Paired patient uncertainty · p. 5, Table II · p. 5

B. Scope of the comparison

  • EF leads both implemented comparators in external point estimates, but capacity, two-view exposure, wall time, and family search effort are unequal.The EF objectives share capacity but are not node-only or parameter-matched architectural ablations; the measurements do not establish a universal clinical ranking.Source: B. Scope of the comparison · p. 5

A. Measured independent evidence

  • Across a fixed 160-image audit, maximum signed-decomposition and prediction-logit errors were 8.76e-06 and 5.96e-06; all 2,834 reachable routes matched.Each class covered 9,216 pixels; mean aligned signed distance was 0.176 and mean typed-cover disagreement was 0.878. This was a mathematical audit, not a diagnostic-accuracy sample or all-test-patch reconstruction.Source: A. Measured independent evidence · p. 5

B. Additional outputs and verification

Evidence Field returns spatial covering-program records that expose the signed computation behind its logits and can be independently checked. The aligned-stability materials distinguish program disagreement from localization or guaranteed stability and identify analyses that remain unavailable.

B. Additional outputs and verification

  • EF records structural signatures, selected scale and node/pair types, five signed non-bias totals, a once-only cover, and an authenticated reconstruction target.An independent reader regenerates descriptors, typed scores, reachable routes, and logits from the input and source-checkpoint coefficients.Source: B. Additional outputs and verification · p. 5

C. Aligned stability

  • 0.9420 and 0.9494 were the held-out AUROCs for three class-centred-loss fits, versus 0.9432 and 0.9451 for the signed variant; paired patient-bootstrap differences included zero at both centres.The headline paired-cross-entropy model has no evidence-consistency term, and these fits do not replace its ensemble or retrospectively change its training loss.Source: C. Aligned stability · p. 5
  • The fixed audit examples show selected class-1 supports and signed contributions whose component differences sum to the logit margin, but these computational supports do not establish lesion localization.The lower tumour-positive example is missed; field scales are panel-specific.Source: Figure 4 · p. 6
  • Variant-specific aligned-stability and full-native-replay measurements remain unavailable, and matched paired-view baseline fits were omitted from the shortened validation study.Source: C. Aligned stability · p. 6

D. Ensemble replay

The ensemble replay audits member programs and raw and calibrated probability mixtures, while the pathology-review output returns probabilities, confidence, spatial components, and a checkpoint-bound program. The paper also provides a reproducibility record, but computational reconstruction is not diagnostic validation.

D. Ensemble replay

  • Probability-quality reporting gives ECE before and after source-temperature calibration, while Brier score and NLL use calibrated ensembles.Source: Table V · p. 7
  • Source retention targets of 75/90% use confidence cutoffs of 0.9564/0.8253; reported retention is the achieved external coverage.Retained error is conditional on accepted patches, with 95% patient-cluster intervals.Source: Table VI · p. 7
  • 2.81×10−7 maximum error reproduced raw and calibrated probability mixtures in a fixed 340-image, three-member audit under three rigid views.The audit covered 6,120 class programs; classification estimates instead use all released test patches.Source: D. Ensemble replay · pp. 6–7

E. Pathology review output

  • An incoming patch returns class probabilities, source-calibrated confidence, region bounds and types, signed components, and a checkpoint-bound program.Institutions can check provenance and replay before applying a locally validated review policy; clinical usefulness and time savings require prospective evaluation.Source: E. Pathology review output · p. 7

VIII. CONCLUSION AND REPRODUCIBILITY

  • The reproducibility record retains preparation, 18 checkpoints and histories, source policies, 54 prediction receipts, statistics, frozen source, and independently reconstructed programs.Code, an independent reader, configurations, a reproducible input, model states, and hash-bound artifacts are available.Source: §VIII · p. 7
Loading 2610.21115v1…