Li, Hao; Jiang, Nan. (2026).听.听The Visual Computer, 42(9), 383.听
贬测产谤颈诲听convolutional neural network (CNN)听补苍诲听Transformer聽models have become popular for analyzing 3D medical images because they combine the strengths of both approaches: CNNs are effective at identifying local image features, while Transformers capture broader patterns across an image. This study introduces聽DSR (Decoder Supervision and Reconstruction), a new training strategy designed to improve how these models learn to reconstruct detailed 3D segmentations without changing how they operate once deployed. During training, DSR adds temporary components that provide extra supervision at multiple stages of the model and encourages the final image representation to remain closely aligned with the original scan. These additional components are removed before the model is used, meaning the deployed model remains the same size and complexity. The approach was evaluated on three medical imaging tasks involving abdominal聽CT听蝉肠补苍蝉,听MRI聽scans of聽vestibular schwannomas聽(noncancerous tumors of the nerve connecting the ear and brain), and prostate MRI. Across all three tasks, DSR achieved the highest聽Dice scores聽(a measure of how closely the predicted segmentation matches the true anatomy) compared with the methods evaluated and also improved measures of boundary accuracy. These findings suggest that strengthening the training of the decoder鈥攖he part of the model that reconstructs the final segmented image鈥攃an improve the accuracy of 3D medical image segmentation without increasing the computational demands of the final model.
Fig 1
Overview of DSR-Net. The input volume is encoded by parallel 3D CNN and Swin Transformer encoders, whose multi-scale features are fused and passed to a shared convolutional decoder. In the decoder,聽聽initializes the first decoder block, while聽,聽,聽, and聽聽are fused into the subsequent stages that produce聽,聽,聽, and聽, respectively. DSR adds training-time decoder branches: auxiliary segmentation heads supervise intermediate decoder features, and an image reconstruction head regularizes the highest-resolution decoder feature. These auxiliary branches are removed at inference, leaving only the main encoder, decoder, and final segmentation head