Organ masks applied in feature space improve weakly supervised scan-level CT classification

Area of Science:

• Medical Imaging
• Artificial Intelligence in Medicine

Background:

• Automated analysis of computed tomography (CT) scans is crucial for medical imaging research.
• Current classification methods often lack spatial information, and using pretrained models with input-space cropping can degrade performance by altering the input distribution.

Purpose of the Study:

• To investigate the efficacy of organ masks in providing scalable anatomical guidance for weakly supervised classification using a frozen self-supervised Swin Transformer.
• To compare input-space organ centering with feature-space cropping for CT scan classification.

Main Methods:

• Two strategies were evaluated: input-space organ centering and feature-space cropping, which applies organ masks to intermediate feature maps before pooling.
• The study utilized three datasets and seven binary classification tasks with a frozen self-supervised Swin Transformer.

Main Results:

• Feature-space cropping matched or improved performance compared to full-image baselines across datasets and tasks.
• Feature-space cropping achieved a pooled improvement of 0.018 in the area under the receiver operating characteristic curve, with significant gains in liver lesion and pericardial effusion classification.
• Input-space centering showed task-dependent effects, and tighter crops in ablation experiments reduced performance, emphasizing the importance of peri-organ context.

Conclusions:

• Feature-level anatomical guidance via feature-space cropping is an efficient strategy to enhance weakly supervised CT classification without retraining the encoder.
• Feature-space cropping reduces embedding dimensionality without performance loss and preserves input distribution.
• This approach offers a scalable method for improving anatomical focus in medical image analysis.