AI-usage in Medical Imagery and Drug Discovery: Strengths and Limitations

Developing a new drug costs an average of 2.6 and takes over 10 years, with less than 10% of candidates reaching patients (Deng et al., 2021). At the same time, radiologists, or medical imaging specialists, work hard to find diagnosis and information from scans. Artificial intelligence, or AI has been growing and advancing over the years, possibly holding the key to faster, cheaper, and more successful drug development and medical diagnosis for patients. However, it is important to consider just how well AI works in medical settings and where it can fall short. 

Artificial intelligence, commonly abbreviated as AI, is essentially a computer program that runs methods and calculations to act as human intelligence. In modern times, one of the most important components of AI is deep learning, which takes raw data and learns from it directly. The average person has probably encountered AI in everyday life when using virtual assistants, seeing Google automatically fill the search bar, or small questions asked to ChatGPT.

However, these types of AI are large LLMS, or large language models. In healthcare, the AI used is very different from the models most are used to using online. Medical imaging uses convolutional neural networks, or CNNs (Sarvamangala & Kulkarni, 2021).These are a type of deep learning process but work by finding disease specific patterns among hundreds of images already verified by professionals.

Drug discovery, on the other hand, may use AI to predict molecules and their functions or reactivity with other molecules. These AIs are trained on hundreds of molecules already verified by scientists to find similar behavior. Protein prediction AI models, with AlphaFold being very popular, predict protein structures and amino acid sequences. The AI may even propose completely new molecules entirely (Vijayan et al., 2021). Since these are not LLMs there are no “hallucinations,” as many are familiar with. Although these AI models might propose an incorrect diagnosis, molecule, or amino acid sequence, they do not create randomly new information (Öğüt, 2025). 

CNN AI models can be helpful to both providers and patients in diagnosis and efficiency. A meta analysis of 279 imaging studies found that AIs using deep learning have sensitivity, which is the ability to correctly identify true positive cases, and specificity, which is the ability to correctly identify true negative cases, comparable to healthcare professionals in selected tasks (Aggarwal et al., 2021). For example, the CNNs detected skin cancer with AUC scores of 0.94 to 0.96, which matched with professional dermatologists. An AUC score is a range from 0 to 1 that measures diagnostic accuracy, with a higher score meaning better accuracy (Guo et al., 2024).

The first FDA-approved AI for diagnosis is the IDx-DR (FDA, 2018) and screens for diabatic retinopathy without a physician. AI can also help with efficiency by running a full analysis and writing reports. With AI assistance radiologist scores rose significantly according to a study of 111 radiologists (Guo et al., 2024). Even with low access to digital resources, CNNs were able to read medical scans quickly. Further, newer AI models can read multiple scan types, including MRIs, PET and CT scans, and compile them into one analyses, for example in diagnostic platforms that use multiple scans (Khened et al., 2021). This can help radiologists and healthcare providers to identify problems sooner.

AI can have also proven to be helpful in drug discovery in producing higher quality predictions and, again, efficiency.

One of the most challenging parts of drug discovery is identifying where to begin and what to experiment with. AI can flag likely candidates by predicting behavior and properties, such as reactivity, absorption, toxicity, and more (Öğüt, 2025), allowing researchers to focus more on the quality of experimentation and trials to push out new drugs to patients sooner. AlphaFold, a protein prediction AI also helps generate structures, which can then be used to determine functions. A recent example of AI’s success in drug discovery actually comes from tweaking an existing drug. An antibiotic against drug resisting pathogens was found within weeks with the help of AI (Nizhenkovska et al., 2024). Neural AI networks scan through millions of compounds within a matter of seconds, making it so efficient while also reducing the cost to scour through all of them.

Another successful example is Insilico Medicine, where a kinase inhibitor was designed in just 46 days and then ready for clinical trials (Öğüt, 2025). Aside from identifying existing compounds, generative AI models can create entirely new molecular structures, not yet listed in the database library of millions of molecules, which can then be tested and verified, effectively expanding the amount of materials we have to work with. 

Despite all the technical advances and benefits of AI, it comes with limitations. First of all, CNNs and neural networks for medical imaging and drug discovery, both suffer from the “black-box” problem. This is when the internal decision making system of the AI is unclear and hard to interpret, making it hard to verify how its predictions were made (Qiu et al., 2023). This is not the same as hallucination where AI comes up with non-existent information. 

Furthermore, the lack of diverse data may become an issue. As AI starts to create and add more and more molecules to the library, eventually they will all be too similar in their chemical scaffolds as they were trained on those limited scaffolds. Similarly, most models are trained at one institution or on one database, which can cause problems for differing demographics. To illustrate, IBM’s Watson for Oncology performed worse for Chinese patients since it was only trained on Western data (Giovanola & Tiribelli, 2022), and the chest X-ray imaging disparities ranged across differing sexes, ages, races, and even insurance type. 

AI models in the sciences and healthcare are growing and shifting at a rapid pace, moving from generalizations to more specific information, analyzing both images and text. It has proven to be useful through increasing diagnostic accuracy and improving efficiency in many areas. However, it still faces limitations with the black-box problem, incorrect generalizations across different demographics, and having a biased algorithm.

In order to make advances in healthcare while still keeping ethical boundaries, regulations on AI privacy and proper unbiased training of models is important. AI is not a replacement for human checks and testing, but rather a supplement. AI usage in medicine should be transparent and clinically valuable for all of us.