Recent advances in machine learning, specifically deep learning, are poised to transform imaging focused healthcare domains including radiology and ophthalmology. But Artificial Intelligence (AI) in medical imaging can be a double-edged sword. Using examples from radiology, oncology, and ophthalmology, we will consider many of the challenges and pitfalls including silent failure of models, lack of repeatability and reproducibility, brittleness, model aging, and a lack of explainability. Models have the potential to entrench and propagate biases. Achieving fairness for all populations is continual challenge. Following a discussion of bias, ethics, fairness and access to equitable healthcare, we will consider some strategies for mitigation including methods for uncertainty estimation and explainability, continuous learning, federated learning, and adversarial training. We will discuss the use of Bayesian deep learning, conformal sets, uncertainty estimation and other technical approaches for increasing model performance, improving fairness and reducing bias. We will consider the ethical dilemmas as we balance increasing access to healthcare through the use of AI, with the challenges of minimizing harm to vulnerable populations. We will also review how medical imaging AI can be used as a lens to study populations and healthcare systems.

The purpose of this study is to devise a Computer Aided Diagnosis (CAD) system that is able to detect COVID-19 abnormalities from chest radio-graphs with increased efficiency and accuracy. We investigate a novel deep learning based ensemble model to classify the category of pneumonia from chest X-ray images. We use a labeled image dataset provided by Society for Imaging Informatics in Medicine for a kaggle competition that contains chest radio-graphs. And the task of our proposed CAD is to categorize between negative for pneumonia or typical, indeterminate, atypical for COVID-19. The training set (with labels publicly available) of this dataset contains 6334 images belonging to 4 classes. Furthermore, we experiment on the efficacy of our proposed ensemble method. Accordingly, we perform a ablation study to confirm that our proposed pipeline drives the classification accuracy higher and also compare our ensemble technique with the existing ones quantitatively and qualitatively.

Opportunistic disease detection on low-dose CT (LDCT) scans is desirable due to expanded use of LDCT scans for lung cancer screening. In this study, a machine learning paradigm called multiple instance learning (MIL) is investigated for emphysema detection in LDCT scans. The top performing method was able to achieve an area under the ROC curve of 0.93 +/- 0.04 in the task of detecting emphysema in the LDCT scans through a combination of MIL and transfer learning. These results suggest that there is strong potential for the use of MIL in automatic, opportunistic LDCT scan assessment.