Industry
AI & ML
Market
Canada
The Probabilistic Vision Group (affiliated with MILA) is an internationally-recognized, interdisciplinary research lab focused on developing probabilistic machine learning frameworks in computer vision developed for a wide range of real-world applications in neurology and neurosurgery.
Challenges
Overcoming limitations in MS lesion detection
PVG faced a significant challenge: improving the accuracy and efficiency of identifying new and enlarging lesions in Multiple Sclerosis (MS) patients. Traditional diagnosis processes, heavily reliant on manual analysis by experts, were not only tedious and time-consuming but also prone to errors and inconsistencies. Moreover, deep learning methods require substantial data, and clinical datasets are often limited, creating further constraints for developing highly accurate diagnostic tools. PVG needed a robust, accurate, and scalable solution that could enhance prediction accuracy despite limited clinical datasets, reduce manual errors, and be adaptable to varying MRI protocols and scanner hardware.
Leveraging transfer learning and hyperparameter tuning for efficiency
Neural Lab addressed these challenges by developing a specialized deep learning model leveraging transfer learning and advanced hyperparameter tuning. We used a large, diverse in-house dataset to initially train our deep learning architecture, known as nnU-Net. This pre-training significantly improved model performance on smaller clinical datasets, such as the MSSEG-2 dataset. Our approach involved fine-tuning the pre-trained model specifically for the MSSEG-2 dataset, capturing intricate patterns necessary for precise lesion detection. Additionally, hyperparameter tuning was conducted to optimize model accuracy further, ensuring adaptability to the specific conditions and imaging protocols used by various clinics. The model took FLAIR MRI scans from two different time points as inputs and generated precise segmentation maps indicating new and enlarging MS lesions. This automation significantly reduced the potential for human bias, provided consistent labeling accuracy, and minimized the resources needed for manual segmentation.
Achieving high accuracy and improved clinical efficiency
The implementation of Neural Lab’s deep learning model led to substantial improvements to PVG's scores (from previous models). Our fine-tuned model achieved an impressive 94% success rate in accurately predicting new and enlarging lesions, marked by an improved F1-score from 0.516 to 0.662 after applying transfer learning and hyperparameter tuning.