Recent advances in machine learning have been actively leveraged for improvements in the medical field. Despite performing very well in a variety of tasks, these models often fail to clearly understand how certain visual changes affect machine learning decisions. Although these AI models have shown great potential and even human capabilities in some cases, there is still a high demand for explanations of what signals these models have learned. Such explanations are essential to build trust among medical professionals and discover new scientific insights from data not yet recognized by experts. Focusing specifically on the lack of explainability in AI models, Google researchers have presented a new framework, StylEx, that leverages generative AI to address challenges in the field of medical imaging.
Current methods for explaining AI models in computer vision, especially medical imaging, often rely on techniques that generate heatmaps that show the importance of different pixels in an image. These methods are useful for showing the “where” of important features, but fall short in explaining the “what” and “why” behind these features. Specifically, they typically do not explain high-level characteristics such as texture, shape, or size that may underlie the model's decisions. To overcome these limitations, Google's StylEx leverages a StyleGAN-based image generator guided by a classifier. The approach aims to generate hypotheses by identifying and visualizing visual signals that correlate with the classifier's predictions.
The workflow involves four main steps: training a classifier to ensure that relevant signals are present in the images, training a StylEx model to generate images following the classifier's instructions, automatically detecting and visualizing key visual attributes that influence the classifier, and reviewing these results with a multidisciplinary expert panel to generate hypotheses for future research. First, the proposed workflow starts with training a classifier on a specific medical image dataset to perform a specific task, ensuring that the classifier achieves high performance (accuracy above 0.8). This step ensures that the images contain information relevant to the task.
We then train a StyleGAN2-based generator to generate realistic images while preserving the classifier's decision-making process. The generator is tuned to focus on attributes that significantly affect the classifier's output. The third stage automatically selects the top attributes in the generator's StyleSpace that influence the classifier's predictions. For each image, the researchers manipulate each coordinate in the StyleSpace to measure its impact on the classification output and identify attributes that significantly change the predictions. This process results in a counterfactual visualization where each attribute is tuned independently, showing its impact.
Finally, a multidisciplinary panel of experts, including clinicians, social scientists, and machine learning engineers, reviews these visualizations. The panel interprets the attributes and determines whether they correspond to known clinical features, potential biases, or novel findings. The panel's insights are then used to generate hypotheses for further research that consider both biological and socio-cultural determinants of health.
In conclusion, the proposed framework increases the explainability of AI models in medical imaging. The approach allows for a deeper understanding of the “what” behind the model's decisions by generating counterfactual images and visualizing attributes that influence the classifier's predictions. The involvement of a multidisciplinary panel beyond physiology and pathophysiology ensures that these insights are rigorously interpreted, potential biases are taken into account, and new avenues of scientific investigation are suggested.
Please check Papers and blogs. All credit for this work goes to the researchers of this project. Also, don't forget to follow us: twitter. participate Telegram Channel, Discord Channeland LinkedIn GroupsUp.
If you like our work, you will love our Newsletter..
Please join us 44k+ ML Subreddit
Pragati Jhunjhunwala is a Consulting Intern at MarktechPost. She is currently pursuing her B.Tech from Indian Institute of Technology (IIT) Kharagpur. She is a technology enthusiast with a keen interest in the range of applications of software and data science. She is constantly reading about developments in various areas of AI and ML.
