IDT: An incremental deep tree framework for biological image classification.

Area of Science:

• Artificial Intelligence in Medical Imaging
• Machine Learning for Disease Detection
• Computational Pathology

Background:

• Breast and cervical cancers are leading causes of cancer death in women.
• Automated AI systems can improve early cancer diagnosis, treatment, and survival rates.
• Convolutional Neural Networks (CNNs) face challenges with catastrophic forgetting when learning new medical image classes.

Purpose of the Study:

• To propose an Incremental Deep Tree (IDT) framework for biological image classification.
• To overcome the catastrophic forgetting problem in CNNs for incremental learning.
• To enable AI models to learn new classes while retaining accuracy on previously learned ones.

Main Methods:

• Development of the Incremental Deep Tree (IDT) framework.
• Evaluation of IDT against established incremental learning methods (iCaRL, LwF, SupportNet).
• Testing on diverse datasets including MNIST, BreakHis, LBC, and SIPaKMeD.

Main Results:

• The IDT framework achieved 87% accuracy on the MNIST dataset.
• Promising results were obtained on medical image datasets: 92% (BreakHis), 98% (LBC), and 93% (SIPaKMeD).
• IDT demonstrated effective incremental learning capabilities, mitigating catastrophic forgetting.

Conclusions:

• The proposed IDT framework offers a viable solution to catastrophic forgetting in deep learning for medical image analysis.
• IDT facilitates continuous learning of new diagnostic classes without compromising performance on existing ones.
• This approach holds significant potential for advancing AI-driven early cancer detection systems.