Liver CT diagnosis switches to composing findings: diseases unseen in training are detected and localized
Synopsis
Liver CT diagnosis models usually predict disease labels directly from images, so they only recognize diseases seen in training. DbC splits diagnosis into two steps: it learns a set of radiological findings as dense maps over the scan, then defines each disease by the findings its guidelines require and composes the diagnosis map with a fixed, parameter-free soft-minimum operator. On multi-phase liver CT, for six diseases never seen in training, DbC localized lesions significantly better than every comparator, including models that align images with disease names, and kept its advantage on rare diseases from another institution and on a public external cohort.
Interpretation
Liver CT diagnosis no longer treats the disease label as the unit of learning; the model first learns a set of nameable radiological findings and then reads a disease as a combination of them, so a disease never seen in training can still be detected and localized. Earlier models predicted disease labels directly from images, with outputs confined to the diseases they were trained on, and the visual findings either never entered the model or were recovered afterwards; vision-language aligned models can be prompted with a disease name at test time but still align images with disease names rather than findings. Under leave-one-disease-out, with each disease removed from training and validation in turn, DbC localized all six unseen diseases significantly better than every comparator and detected four of the six significantly better; the disease-name-aligned ablation Disease-VL, sharing encoder, data and compute, detected FNH at an AUROC of .532, close to chance.
The model trains a single task, one probability map per finding over the whole scan for 40 findings, and a disease diagnosis map is the token-wise soft minimum of its required finding maps, an operator with no learned parameters. Each finding map comes from matching image features at every spatial location to the text embedding of the finding name from a frozen BiomedCLIP text encoder, rescaled by a learned per-finding scale and bias and passed through a sigmoid, the way a vision-language aligned model recognizes a category, but supervised densely rather than by image-text pairs. Training used 15,609 multi-phase liver CTs from a single institution, split at the patient level into 12,487 training, 1,559 validation and 1,563 test CTs; finding labels came from guideline-derived required findings, supportive findings extracted from reports by Qwen3-Max, and size findings derived from lesion masks.
Perspective
The results are limited to liver disease on 3D CT, with training data from a single institution and seven diagnoses; the rare-disease and external-cohort experiments test how the finding vocabulary behaves on diseases it was not designed for. The authors propose that the framework is not specific to the liver and could extend to other organs through a broader catalog of named findings and disease compositions, but that extension is not validated here.
The compositions are derived by experts from published guidelines and require all of their findings, so each lesion's finding labels are only as complete as the guideline and report sources; the authors note that richer composition rules could follow atypical presentations more closely. The authors also describe the reading of the results as a reusable finding vocabulary generalizing better than a set of disease labels as a hypothesized interpretation, and state that the model is a research prototype requiring prospective validation and regulatory clearance before clinical use.