Skip to main content
Back to timeline
arXivSource publication:

Prostate biopsy grading now shows its work: 69 named sparse features make gland architecture readable, removable and testable

Synopsis

Gleason grading of prostate biopsies usually returns a grade without the glandular architecture behind it. This work trains a sparse autoencoder on tokens of the pathology foundation model UNI2-h and aligns it to multi-reader annotations, naming 69 of its 2048 features after morphologies such as cribriform glands, cords and comedonecrosis; the same code feeds a grading network that predicts Gleason patterns and Grade Group and also yields morphology maps directly. Without retraining, cribriform detection reaches AUROC 0.849 on external biopsies and pattern 5 reaches 0.846 to 0.919 against six pathologists on unseen cores, while removing a pattern's named features cuts its predictions by 49% to 98%.

Interpretation

The grading system returns not only a grade but glandular morphology evidence drawn from the same representation: one sparse code feeds the grading network that computes Gleason patterns and Grade Group and also yields six morphology maps directly. Earlier deep learning grading systems reached pathologist-level agreement on external validation but relied on learned computations that obscure how inputs become a diagnosis, while concept bottleneck models route prediction through predefined concepts alone. On the PANDA-Radboud test set the grading network reaches a macro Dice of 0.730, the single sparse U-Net 0.728 and the embedding U-Net that reads UNI2-h tokens directly 0.742, a paired difference of −0.014.

Morphology maps need no separately trained segmentation network: the activations of a morphology's feature set are scaled by fixed statistics and summed at each tissue token. The named sets come from a gated linear layer aligned on 812 GleasonXAI tissue-microarray cores against three readers' annotations, and a feature joins a class when its gate for that class ends above 0.5. On GleasonXAI, with each map taken from an alignment that excluded the core's reader panel, token-level AUROC is 0.949 for cords, 0.943 for solid groups, 0.921 for well-formed glands and 0.899 for cribriform glands, 0.759 for poorly formed and fused glands and 0.719 for comedonecrosis.

The contribution of named features to grading is measured directly by setting a pattern's named features to zero before the code enters the grading network, leaving all other features and network parameters unchanged. As a control, 20 sets of unnamed features of the same size and matched activation frequency are removed in the same way for each pattern. On PANDA-Radboud test slides, removing the pattern 3, 4 and 5 named sets reduces the corresponding pattern calls by 97.6%, 48.7% and 89.9%; on 100 test slides the random sets change them by −9.5% to +11.1%, −8.3% to +11.9% and −14.9% to +31.1%, and each named removal ranks first among the 21 removals.

Perspective

The demonstrated scope is five public H&E prostate datasets: PANDA-Radboud for pretraining and grading-network training and selection, GleasonXAI for alignment, and Gleason2019, SICAPv2 and PAR for evaluation only, covering needle biopsies, tissue-microarray cores, image patches and three scanners. The intended reader is a computational pathology or medical imaging researcher; the system uses settings unchanged on the external cohorts.

Three of the nine annotated morphologies received no feature set because the public annotations contain at most 1362 tokens marked by two readers for each, and comedonecrosis, with only five features, is the least stable map across runs; the solid-group and comedonecrosis sets overlap heavily with cords, mirroring their co-occurrence in the readers' drawings. Grading agreement falls on the Hamamatsu scanner, yet the code keeps its sparsity and reconstruction error on those scans and the embedding U-Net is affected in more comparisons. Evaluation rests on one encoder and retrospective public datasets, so other encoders and prospective use remain to be tested.