Principal Investigator: Dr. Faisal Mahmood
Currently, there exists great disparity in healthcare quality around the world, with many developing areas facing challenges such as lack of pathologists or adequate facilities. Recent advances in the area show that artificial intelligence (AI) could serve as a potential solution to address the shortage of pathologists and equipment. However, the actual deployment and use of AI in diagnostic contexts requires expensive hardware for imaging samples and computing, which is another great challenge in low-resource conditions. To address this concern, we present a cost-effective (less than $500) and easy-to-use device that integrates microscope imaging and AI for diagnosis on whole tissue slides in a streamlined process. The design uses 3D printing and conventional optics and computing parts to assemble the bright-field microscope for imaging tissue slides and to run the AI model. Our AI is designed such that aside from outputting a diagnosis, it also creates a heatmap that shows which regions of the tissue is most influential to its diagnosis, offering human interpretability of the results. We measured the imaging quality and model performance and found that the device was able to achieve high resolution and contrasts and accurately diagnose pathology slides while maintain an affordable and lightweight setup.
The lack of widespread, effective diagnosis in developing areas is a major factor contributing to the disparity in healthcare around the world. Low-resource settings face challenges such as lack of pathologists and adequate facilities. Advances in artificial intelligence models for pathology diagnosis provide viable solutions to address this disparity. However, deployment of these AI models requires expensive hardware for imaging and computing, creating a challenge for low-resource applications. We present a cost-effective device that integrates microscope imaging and deep learning diagnosis of whole pathology slides in a streamlined process. The model was trained from digitized H&E whole slide images with slide-level ground truth labels using attention-based learning. For interpretability, heatmaps of the attention scores are generated to show the regions influential to the model’s classification. To assess functionality and accuracy, we evaluated microscope imaging quality and model performance. For imaging, we measured resolution, color transmission, and distortion, obtaining results comparable to a conventional microscope. For performance, we imaged tissue sections using the setup and assessed the model on several classification tasks. The device was able to image with high resolution and contrasts and accurately classify pathology slides from both FFPE and frozen sections while maintaining an affordable and lightweight setup.