Medical Imaging and Machine Learning

Introduction

Deploying AI systems for clinical tasks is challenging due to issues like input quality control and integration into clinical workflows. Early attempts at clinical machine learning have shown that identifying deployment problems must start at the design phase. Future AI systems need to address systemic biases, uncertainty in predictions, and provide explanations for their decisions.

High Dimensional Medical Imaging Data

The demand for high-quality annotated medical datasets is expected to exceed supply. Creating multicenter imaging datasets requires standardized protocols and addresses privacy concerns. The DICOM standard is crucial for managing medical images, and various tools exist for de-identifying sensitive data.

Computational Architectures

Neural network architectures for clinical machine learning are adapted from those used in image recognition tasks. These architectures can handle high-dimensional data, such as echocardiograms and MRI scans, effectively. Video-based neural networks have shown improvements over traditional 2D approaches.

Time to event analyses and Uncertainty Quantification

As AI systems evolve from diagnostic to prognostic applications, time-to-event analyses become more relevant. These analyses can predict event probabilities over time and integrate with traditional survival models, enhancing the prognostic value of medical imaging.

Explainable AI and Risk of Harm

Saliency maps are commonly used to explain ML predictions, but they can be misleading. Methods that account for uncertainty in predictions are essential to alert clinicians to potential errors. A more granular approach, such as serial occlusion tests, can help validate the features identified by ML systems.

Figures Explained

The paper’s visual material highlights the workflow and the main system components.

• Figure 1 .: Figure 1. Cloud based tools such as MD.ai can be used for generating expert annotated datasets and evaluating against clinical experts via a secure connection.
• Figure 2 :: Figure 2: Modular approach to structuring machine learning code as described by the PyTorch Lightning team. Such frameworks enable rapid iteration and experiment management under a standardized framework, and bridges engineering best practices with the.
• Figure 4a .: Figure 4a. Experiments conducted by Adebayo et. al. (reproduced with permission) with models trained on true labels from the MNIST dataset, and trained on random noise. Models trained on random noise still yield the circular shape of the digit zero when evaluated by the majority of visualization methods. These offer little in terms of true saliency maps, functioning more as class invariant edge detectors. b. Detection of echocardiographic view planes: both incorrect (left) and correct classifications (right) yield similar saliency maps.
• Figure: Figure3. Machine learning models trained with standard methods can be extremely confident even when incorrect, as described by Sensoy et. al (reproduced with permission). As a digit is rotated 90 degrees, the system confidently assigns the label from ‘1’ to ‘7’. On the other hand, with methods that account for classification uncertainty, the system assigns an ‘uncertainty score’ that can help alert clinicians to potentially erroneous predictions.