AUGUST

A pathologist reading a gastric biopsy never arrives at "poorly cohesive carcinoma" in a single step. They establish that a lesion is present, then that it is malignant, then that it is a carcinoma rather than a lymphoma, and only then grade its differentiation — and that last question is meaningless if the first answer was chronic gastritis. Most computational pathology systems ignore that structure, attaching an independent head per label, so nothing stops a slide from being called negative for cancer and graded as poorly differentiated adenocarcinoma.

AUGUST (Adaptive Unified Gastric diagnosis Using Sequential Tasks) makes the dependency explicit by turning diagnosis into a sequential question-answering process. Given a whole-slide image, it generates the next question conditioned on what has already been answered, projects the slide into a representation conditioned on that question, and emits an answer drawn only from the options the clinical taxonomy allows at that point. The output is not a label vector but a transcript, and a single unified checkpoint covers the whole hierarchy rather than one model per task.

The model comes from Jin Tae Kwak's group in the School of Electrical Engineering at Korea University, working with hospital pathologists at The Catholic University of Korea, and was published in npj Digital Medicine in August 2026. Code and gated weights were released earlier, in January 2026.

#Key Features

• Conditional question generation: Each step's question is constructed from the answers to prior steps, so the model narrows from diagnostic category to subtype to grade instead of predicting all levels at once.
• Answers constrained to valid pathways: Candidate answers at every turn are restricted to the options the taxonomy permits given the upstream findings, which is what removes self-contradictory outputs and raises hierarchical consistency.
• Question-conditioned slide encoding: The question embedding is appended to the patch sequence before aggregation, so the slide representation the language model sees is built for the question being asked rather than reused unchanged across tasks.
• Human-in-the-loop operation: A pathologist can supply the answer to an intermediate question at inference time; the model continues the chain from that answer. This is a correction at inference, not retraining.
• Auditable reasoning trace: Because the diagnosis is assembled from a readable series of questions and answers, a reviewer can see which step a disagreement originates in.

#Technical Details

AUGUST couples a slide encoder to a frozen 8-billion-parameter medical language model (Med-LLaMA3-8B), adapted with rank-16 LoRA on the attention and feed-forward projections during the later training stages. Slides are tiled at 512×512 pixels at 20× and encoded into 1024-dimensional patch features with UNI; an S4 state-space multiple-instance-learning encoder aggregates that sequence — with the question embedding inserted as an additional token — into a single slide vector, which a fusion adapter concatenates with the question embedding before handing it to the language model. Auxiliary classifier heads for stomach location, Helicobacter pylori status and coarse diagnostic category supervise the slide representation directly. Training runs in three sessions: cross-domain alignment on 6,913 WSI–caption pairs, visual instruction tuning on 27,069 question–answer pairs, then multi-turn VQA that restructures the same material into chain-of-thought sequences.

The training cohort is GastUJB, 6,913 slides collected at Uijeongbu St. Mary's Hospital between 2014 and 2023, split at patient level against a 5,166-slide test set (12,079 slides in total). Evaluation spans 23,072 gastric whole-slide images, 19 hierarchical tasks — three coarse, eleven fine, five grading — and 79,924 question–answer pairs; that slide count is the full evaluation corpus and includes the training cohort rather than being a held-out set. The authors report higher accuracy and hierarchical consistency than multiple-instance-learning, pathology vision-language and pathology foundation-model baselines in both autonomous and human-in-the-loop settings.

#Applications

The task list maps onto routine gastric biopsy sign-out: H. pylori status, Sydney System inflammation grading, benign tumour and dysplasia subtyping, carcinoma versus lymphoma, and adenocarcinoma differentiation. For a pathology service, the draw is the transcript — a second reader that shows its route rather than a bare label, and that can be interrupted mid-chain when a specialist already knows an intermediate answer. The released code also includes the question templates and the pipeline that turns tabular clinical metadata into hierarchical conversations, which groups building assistants for other organs can reuse.

#Impact

AUGUST's contribution is architectural rather than a leaderboard position — it shows that encoding the conditional structure of a diagnostic protocol into the model, rather than recovering it post hoc from independent predictions, buys both consistency and interpretability. The limits are real: the taxonomy is fixed and gastric-specific, training data comes from a single Korean hospital, and the reported comparisons are the developers' own. Weights are gated behind an access request, the code carries no open-source license, and the authors restrict use to academic research, excluding commercial deployment and clinical decision-making.