An Improved Multi-Instance Learning Model for Postoperative Early Recurrence Prediction of Hepatocellular Carcinoma Using Histopathological Images

HCC(Hepatocellular carcinoma) is one of the primary liver cancer around the world, and it produces a high mortality rate in clinical. Surgical resection is the first-line treatment choice for patients with HCC, but the problem is patients will still have a great chance of recurrence after surgery. Therefore postoperative early recurrence prediction of HCC patients is necessary, which can guide physicians to manage the individualized follow-up so that we could increase the survival time of patients. MIL(multi-instance learning) model is mostly utilized in histopathological image analysis, and it has achieved promising results in most tasks, but it has one major limitation: the conventional MIL model loses the histopathological spatial structure information when generating the slide-level representation on flattened instance embeddings. In this light, we revisit the conventional MIL model and propose an improved MIL model to tackle the limitation. Our proposed method consists of three stages, the first stage is tissue patching, which is the same as the conventional MIL; the second stage is early recurrence probability heatmaps(ER-ProbMaps) generation, which is conducted by a well-trained patch-level prediction model on the tissue patches of the same slide; the final stage is to produce the slide-level recurrence prediction by a slide-level prediction model. We train the slide-level prediction model on obtained heatmaps to complete our postoperative early recurrence prediction task. Our experimental results show that our proposed improved MIL model can achieve superior performance than the conventional MIL model in our task.