Meet ConDistFL: A Revolutionary Federated Learning Approach for Organ and Disease Segmentation in CT Datasets

Computed tomography (CT) images must accurately segment abdominal organs and tumors for clinical applications like computer-aided diagnosis and treatment planning. A generalized model that can handle numerous organs and illnesses simultaneously is preferred in real-world healthcare circumstances. While major research has concentrated on segmenting individual organs and different classes of organs without malignancy, there are other areas of interest. On the other hand, traditional supervised learning techniques rely on the volume and caliber of training data. Unfortunately, a lack of training data resulted from the expensive expense of high-quality medical imaging data. Only qualified specialists can create correct annotations on medical pictures for various anatomies. 

It is also difficult to annotate the organs and associated cancers of diverse anatomies and imaging modalities since even professionals sometimes only have specialized expertise for a single activity. The development of generalized segmentation models is significantly hampered by the need for more suitable annotated information for various organs and malignancies. Numerous research has investigated partly annotated datasets, where only a portion of targeted organs and malignancies are tagged in each picture, to develop generalized segmentation models to solve this issue. However, sharing confidential medical statistics among organizations presents privacy and legal issues. Federated learning (FL) was proposed to address these issues. 

FL enables collaborative training of a common (or “global”) model across several institutions without centralizing the data in one place. A potential method to increase the effectiveness of medical picture segmentation is FL. In FL, each client merely sends model updates to the server and instead uses its data and resources to train a local model. The server uses “FedAvg” to integrate these changes into a global model. Recent research has used FL to create unified multi-organ segmentation models utilizing abdominal datasets that were only partially annotated, as seen in Fig. 1. These methods, nevertheless, frequently ignore lesion regions. Few studies have made an effort to segment the various organs and their tumors at the same time. 

Due to the difficulty in dealing with data heterogeneity caused by data variety, FL’s model aggregation faces significant challenges. Performance might suffer when models from diverse sources are used with non-IID data. When clients use data annotated for various purposes, more domain shifts in the label space are introduced, making the problem worse. Additionally, the performance of the global model for jobs with less data may be impacted by clients’ differing dataset sizes. Researchers from National Taiwan University,  Nagoya University and NVIDIA Corporation in this paper offer a strategy to deal with data heterogeneity in FL for multi-class organ and tumor segmentation from partially annotated abdominal CT images. 

These are the primary contributions of this work: 

1. their proposed conditional distillation federated learning (ConDistFL) framework makes the combined multi-task segmentation of abdominal organs and malignancies possible without the need for additional fully annotated datasets. 

2. In real-world FL settings, the proposed framework shows stability and performance with lengthy local training steps and a small number of aggregations, lowering data traffic and training time. 

3. They use an unreleased, fully annotated public dataset called AMOS22 to test their models further. The qualitative and quantitative evaluations’ findings demonstrate their strategy’s robustness.

Check out the Paper and GitHub. All Credit For This Research Goes To the Researchers on This Project. Also, don’t forget to join our 28k+ ML SubReddit, 40k+ Facebook Community, Discord Channel, and Email Newsletter, where we share the latest AI research news, cool AI projects, and more.