Saibro Güinther, Keeza Yvonne, Sauer Benoît, Marescaux Jacques, Diana Michele, Hostettler Alexandre, Collins Toby
Ircad Africa, Kigali, Rwanda.
Ircad France, Strasbourg, France.
Int J Comput Assist Radiol Surg. 2025 May;20(5):923-933. doi: 10.1007/s11548-025-03334-z. Epub 2025 Mar 11.
Despite major advances in Computer Assisted Diagnosis (CAD), the need for carefully labeled training data remains an important clinical translation barrier. This work aims to overcome this barrier for ultrasound video-based CAD, using video-level classification labels combined with a novel training strategy to improve the generalization performance of state-of-the-art (SOTA) video classifiers.
SOTA video classifiers were trained and evaluated on a novel ultrasound video dataset of liver and kidney pathologies, and they all struggled to generalize, especially for kidney pathologies. A new training strategy is presented, wherein a frame relevance assessor is trained to score the video frames in a video by diagnostic relevance. This is used to automatically generate diagnostically-relevant video clips (DR-Clips), which guide a video classifier during training and inference.
Using DR-Clips with a Video Swin Transformer, we achieved a 0.92 ROC-AUC for kidney pathology detection in videos, compared to 0.72 ROC-AUC with a Swin Transformer and standard video clips. For liver steatosis detection, due to the diffuse nature of the pathology, the Video Swin Transformer, and other video classifiers, performed similarly well, generally exceeding a 0.92 ROC-AUC.
In theory, video classifiers, such as video transformers, should be able to solve ultrasound CAD tasks with video labels. However, in practice, video labels provide weaker supervision compared to image labels, resulting in worse generalization, as demonstrated. The additional frame guidance provided by DR-Clips enhances performance significantly. The results highlight current limits and opportunities to improve frame guidance.
尽管计算机辅助诊断（CAD）取得了重大进展，但对精心标注的训练数据的需求仍然是一个重要的临床转化障碍。这项工作旨在克服基于超声视频的CAD的这一障碍，使用视频级分类标签结合一种新颖的训练策略来提高最先进（SOTA）视频分类器的泛化性能。
在一个关于肝脏和肾脏病变的新型超声视频数据集上对SOTA视频分类器进行训练和评估，它们都难以实现泛化，尤其是对于肾脏病变。提出了一种新的训练策略，其中训练一个帧相关性评估器，以根据诊断相关性对视频中的视频帧进行评分。这用于自动生成与诊断相关的视频片段（DR-片段），在训练和推理过程中指导视频分类器。
使用带有视频Swin Transformer的DR-片段，我们在视频中肾脏病变检测方面实现了0.92的ROC-AUC，相比之下，使用Swin Transformer和标准视频片段时的ROC-AUC为0.72。对于肝脏脂肪变性检测，由于病变的弥漫性，视频Swin Transformer和其他视频分类器表现相似，通常超过0.92的ROC-AUC。
理论上，视频分类器，如视频Transformer，应该能够通过视频标签解决超声CAD任务。然而，在实践中，与图像标签相比，视频标签提供的监督较弱，导致泛化性较差，如所示。DR-片段提供的额外帧指导显著提高了性能。结果突出了当前改进帧指导的局限性和机会。
