Scalable Clinical Annotation with Location Evidence (SCALE)

Area of Science:

• Medical Imaging AI
• Machine Learning in Radiology
• Prostate Cancer Diagnostics

Background:

• Deep learning for medical imaging requires large annotated datasets, which are difficult to obtain.
• The global shortage of radiologists necessitates efficient AI development for medical image analysis.
• Automated annotation methods are crucial for scaling AI development in healthcare.

Purpose of the Study:

• To introduce SCALE (Scalable Clinical Annotation with Location Evidence), a fully automated method for generating voxel-level annotations.
• To develop and train an optimized AI algorithm using large-scale datasets annotated with SCALE.
• To evaluate the performance of AI models trained with SCALE annotations against other methods for prostate cancer detection on MRI.

Main Methods:

• Developed SCALE, a method utilizing location priors from medical reports, biopsy coordinates, or anatomical sectors for automated annotation.
• Annotated a large dataset of 17,896 cases using both SCALE and a count-based weakly semisupervised learning (CWSSL) method.
• Trained and evaluated an optimized AI algorithm on datasets generated by SCALE and CWSSL, comparing performance against supervised learning and the PI-CAI Ensemble AI System.

Main Results:

• The AI model trained on SCALE-annotated data achieved a case-level area under the receiver operating characteristic curve (AUC) of 0.856.
• This performance was superior to models trained with supervised learning (AUC +0.012, p=0.02) and comparable to CWSSL (AUC +0.007, p=0.12).
• The SCALE-trained model also showed a slight advantage over the PI-CAI Ensemble AI System (AUC +0.006).

Conclusions:

• Automated, location-guided annotation with SCALE enables scalable development of AI for clinically significant prostate cancer detection on MRI.
• The SCALE method surpasses previous annotation and AI training approaches, facilitating broader clinical deployment of AI tools.
• This work demonstrates the potential of automated annotation to address data limitations in medical AI research and development.