AI Is a Tool for Radiology, Not a Replacement for the Profession

“AI will overtake radiology” is a claim we have all heard by someone at some point, and it has become even more redundant with recent developments, such as OpenAI’s ChatGPT.

While these conversations boast about the efficacy of AI, they often fail to detail how AI programs are designed for use in radiology and the skill sets that radiologists possess. The proper context for these points may help readers understand why AI will undoubtedly become a pivotal tool for radiologists but is unlikely to replace the profession.

Computer vision allows for the analysis and interpretation of medical images, identifying anomalies like tumors or fractures. Deep learning, particularly through convolutional neural networks (CNN), enhances the accuracy of diagnosing diseases from these images.

The first step involves gathering a substantial dataset of annotated medical images. These annotations, often made by expert radiologists, highlight regions of interest and label abnormalities and unique features.

The data is then preprocessed, which includes normalization to ensure consistent image quality, and augmentation to increase dataset diversity. CNNs are designed for specific diagnostic tasks and operate like a digital detective for images. It starts by spotting small details, like colors and edges, and then combines these clues to recognize larger objects or patterns, such as faces, animals or cars. This model is trained using the prepared dataset, iteratively adjusting its parameters to minimize prediction errors.

Natural language processors (NLP) assist in automating the generation and interpretation of radiology reports, streamlining the workflow for radiologists. NLPs are crafted by gathering a large dataset and training machine learning models to understand medical jargon, context and relationships between terms. These models are repeatedly tested and refined using unseen data until they can reliably generate or interpret radiology reports.

ChatGPT is a type of NLP based on the Transformer architecture, which uses the self-attention mechanism¹. This allows the model to weigh the importance of different words in a sentence relative to a given word. For example, in the sentence, "The lungs are clear. There is no evidence of a mass, consolidation or pleural effusion," the self-attention mechanism might associate higher importance to words like "lung" to understand the context better.

Despite these remarkable advances in AI, a radiologist's expertise goes beyond mere image interpretation. Radiologists bring in years of medical training, contextual knowledge of patient histories and a nuanced understanding of disease presentations and progressions. AI lacks the deep human intuition, experience-driven judgment and interpersonal skills essential in healthcare.

AI can assist by streamlining certain processes and offering insights based on data, but it will only ever be as good as the training data and will be less likely to identify rarer abnormalities. Thus, while AI will play an increasingly significant supporting role, the unique blend of technical expertise and human qualities in radiologists ensures they remain irreplaceable in the foreseeable future.

¹Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A.N., Kaiser, L., & Polosukhin, I. Attention Is All You Need. Available at: https://arxiv.org/abs/1706.03762. Accessed Nov. 13, 2023.