MedicalNet: Tencent's Pre-trained Model for 3D Medical Imaging AI

This article introduces MedicalNet, a Tencent open-source project that provides pre-trained models based on 3D medical imaging big data. Medical imaging AI addresses the global challenge of "patients having difficulty accessing healthcare and doctors being exhausted from diagnosis." With the shortage of medical personnel that cannot be quickly resolved, AI technology can assist medical work by performing population-based screening and improving diagnostic quality.

However, the development of medical imaging AI faces a bottleneck: limited annotated data from medical experts creates a contradiction with data-driven deep learning requirements. To address this, MedicalNet aggregates multiple small medical 3D public datasets across different scenarios to form a large-scale dataset for pre-training models.

The technical challenges addressed include: inconsistent pixel meanings, large range differences, frequent artifacts, low imaging quality, blurred boundaries, low contrast, missing annotations across different data sources, inconsistent resolution for the same tissue, and large scale variations across different tissues.

MedicalNet employs two main solutions: (1) Dataset filtering to identify datasets with common knowledge by creating mini-dataset proxies from each scenario, training small networks, and evaluating prediction quality to determine which datasets to retain; (2) Joint training with spatial and pixel normalization preprocessing. MedicalNet consists of encoder and decoder components, with most parameters concentrated in the encoder to gather more information. The decoder uses multi-task learning to isolate annotations from multiple scenarios. Different skip-connection combinations are used to alleviate gradient vanishing during training.

Experimental results demonstrate that MedicalNet helps networks in small-data 3D medical imaging scenarios converge faster and improve prediction performance. The code is open-sourced at https://github.com/Tencent/MedicalNet.