Domain affiliated distilled knowledge transfer for improved convergence of Ph-negative MPN identifier

Area of Science:

• Medical Imaging
• Computational Biology
• Artificial Intelligence

Background:

• Ph-negative myeloproliferative neoplasms (MPNs) are rare, progressive blood disorders requiring accurate diagnosis.
• Current diagnostic methods for MPNs can be labor-intensive, involving multiple pathological data types.
• Deep learning research for MPNs is limited by data scarcity and common augmentation techniques.

Purpose of the Study:

• To address data scarcity in Ph-negative MPN diagnosis using deep learning.
• To enhance the performance of lightweight models by transferring knowledge from larger datasets.
• To investigate the efficacy of knowledge distillation for rare disease image analysis.

Main Methods:

• A 50-layer ResNet was pre-trained on a large dataset of 327,680 lymph node images.
• Knowledge from the ResNet was distilled into a smaller 4-layer Convolutional Neural Network (CNN).
• The lightweight CNN was initialized with distilled weights and fine-tuned on a small dataset of 300 MPN images.

Main Results:

• The CNN model utilizing distilled knowledge achieved a diagnostic accuracy of 97%.
• A comparable CNN trained from scratch on the same small dataset achieved only 89.67% accuracy.
• Knowledge distillation outperformed data augmentation and manual feature extraction in addressing data scarcity.

Conclusions:

• Transferring distilled knowledge significantly boosts the performance of lightweight models for rare diseases.
• Knowledge distillation is an effective strategy for overcoming data limitations in medical image analysis for MPNs.
• This approach offers improved accuracy and convergence for training on limited Ph-negative MPN datasets.