Annotation-Efficient Deep Learning Model for Pancreatic Cancer Diagnosis and Classification Using CT Images: A...

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Oncology

Background:

• Pancreatic cancer (PC) detection often relies on large annotated datasets for deep learning (DL) models.
• Developing effective DL models for PC detection with limited data remains a challenge.

Purpose of the Study:

• To develop a novel deep learning model for pancreatic cancer detection using computed tomography (CT) images without requiring large annotated training datasets.
• To evaluate the performance of a self-supervised learning algorithm, pseudo-lesion segmentation (PS), in enhancing DL-based PC classification.

Main Methods:

• A retrospective diagnostic study utilized CT images from 4287 patients diagnosed with PC.
• A self-supervised learning algorithm (pseudo-lesion segmentation (PS)) was proposed and integrated into convolutional neural network (CNN) and transformer-based DL models.
• Models were trained with and without PS, validated internally, and externally validated on a separate dataset of 361 patients.

Main Results:

• Internal validation showed high accuracy and sensitivity for both CNN and transformer models with PS (e.g., transformer model achieved 95.7% accuracy and 99.3% sensitivity).
• Implementing PS on a small dataset (10% of training data) significantly increased accuracy by 20.5% and sensitivity by 37.0%.
• External validation demonstrated robust performance, with the transformer model achieving 87.8% accuracy and 86.5% sensitivity.

Conclusions:

• Pseudo-lesion segmentation (PS) self-supervised learning effectively enhances the performance, reliability, and robustness of DL models for pancreatic cancer classification.
• The proposed DL model demonstrates potential utility in PC diagnosis, particularly when dealing with limited or small datasets.