Mitosis-specialised histopathology encoder: a ViT-Base pretrained on 1.8M pseudo-labelled TCGA cell crops to tell mitotic figures from mimickers.
No providers recorded yet. Browse all providers
A pathologist grading a breast carcinoma or a soft-tissue sarcoma counts mitotic figures — cells caught mid-division — in a fixed area of tissue. The hard part is not finding dividing cells but rejecting the look-alikes: apoptotic bodies, pyknotic nuclei and dense lymphocytes that share a dark, condensed appearance at the roughly 32 µm scale where a mitosis is judged. General-purpose pathology encoders such as Virchow and UNI are trained on large tissue tiles sampled without regard to cell type, so the features they learn are not tuned to that fine discrimination, and their size makes them costly to run over the thousands of candidates a whole slide produces.
MiTHras answers this with task-specific pretraining. Instead of generic tiles, it trains a ViT-Base encoder on cell-centred crops that an existing mitosis detector has already sorted into pseudo-mitoses and pseudo-mimickers. Those noisy labels steer a contrastive objective at two scales — crops sharing a label are pulled together, and tokens inside a putative mitosis are separated from the surrounding tissue tokens — while a masked-autoencoder reconstruction term keeps the encoder from discarding general image information. The result is a single pretrained encoder reused across mitosis classification, detection, whole-slide counting and subtyping.
The model was developed by Jin T. Kwak's group at Korea University with Nasir Rajpoot's team at Histofy, together with collaborators at Brigham and Women's Hospital, Harvard Medical School and University Medical Center Ho Chi Minh City. It was released as an arXiv preprint in September 2026.
The pretraining corpus is built in three stages: 896×896-pixel tissue tiles are extracted with Otsu thresholding, tumour regions are kept using the IMPaSh tissue classifier, and a MIDOG22-trained two-stage detector (Efficient-UNet segmentation plus an ImageNet MAE ViT-Base classifier) assigns mitosis and mimicker labels. A pathologist audit put pool-weighted precision of the pseudo-mitoses at 94.9–98.1%, while about 20% of pseudo-mimicker crops contained true mitoses. MiTHras is a standard ViT-Base encoder with 768-dimensional embeddings, initialised from scratch and trained for 200 epochs on eight NVIDIA A6000 GPUs with AdamW at an effective batch size of 512.
Evaluation compares MiTHras with ImageNet encoders, DINOv2, CLIP, PLIP, CTransPath, GPFM, UNI and Virchow over three training folds. On mitosis-versus-mimicker classification it reached F1 scores of 0.864, 0.855 and 0.789 on MIDOG++, a canine tumour set and a held-out TCGA test set under full fine-tuning, and 0.836, 0.848 and 0.775 under linear probing — where the best general-purpose or pathology foundation encoder reached 0.421 on MIDOG++. On the AMi-Br and AMi-Morph subtype benchmarks it scored 0.700 (typical versus atypical) and 0.483 (eight morphology classes), ahead of the strongest baselines by 0.046 and 0.031. Detection gains were modest because every method shares the same candidate segmentation stage. Whole-slide mitotic counts from the pipeline gave C-indices of 0.627 (hotspot) and 0.671 (density) for overall survival on TCGA-GBMLGG.
MiTHras targets automated mitotic counting in grading systems that depend on it, such as Nottingham grading of breast carcinoma, FNCLCC grading of soft-tissue sarcoma and risk stratification of gastrointestinal stromal tumours. Its encoder can serve as the classification stage of a detect-then-classify pipeline over whole slides, as a feature extractor for typical-versus-atypical mitosis subtyping, and as a source of mitotic hotspot and density measurements for prognostic studies.
MiTHras is a data point for task-specific pretraining in computational pathology: a ViT-Base trained on 1.8 million targeted crops outperformed ViT-Large and ViT-Huge encoders trained on far larger generic corpora on the mitosis benchmarks tested, with the widest margins when features were frozen. It complements challenge-driven mitosis classifiers such as EFTD-MIDOG-AMF, which instead adapt a general visual backbone. Its limits are stated in the paper: gains vary across tasks, eight-class morphology remains difficult, the fixed pseudo bounding boxes introduce label noise, and the survival analysis is exploratory, with TCGA-THCA holding only 11 events. No code, weights or the TCGA-MF-Pseudo corpus had been released alongside the preprint.
Much of this page is generated or calculated automatically. Flag anything that looks off and we will re-run it.