Artificial intelligence in medical applications is always faced with a series of difficulties. The challenges range from the quality of input data and model performance to ethical problems.
Some key challenges include:
Poor quality data
Major applications of AI include using digital pathology slides with image analysis algorithms for accurate diagnosis. However, preparing slide images is a cumbersome process involving embedding, cutting, staining, and scanning the tissue sample. Due to this, only a few datasets are available for adequate training.
Moreover, the image preparation process is so complex that the slide samples can become distorted. Elements such as dust, hair, and air bubbles can easily make their way into the slides undetected. This results in poor-quality data and weakly supervised deep-learning models. A weak model presents unreliable results and raises questions about its integrity.
Data dimensionality and hardware limitations
Whole slide imaging deals with data in the size of giga-pixels. A typical slide scan spans around 100,000 by 100,000 pixels in size.
Deep learning algorithms, on the other hand, are often trained using images of around 250 by 250 pixels. This is because the computer hardware used for training cannot store such large-sized files.
As a workaround, images are downscaled before being passed to the model. The images lose vital information during down sampling, resulting in performance degradation.
Lack of clinical and technical expertise
Building an AI model requires experts in clinical-grade computational pathology, statistics, artificial intelligence, and on-field medical professionals. The process consists of the collection and preparation of clinical data, annotation, model training, and, finally, validation.
Each of these experts plays a vital role in the entire process. Medical practitioners guide researchers regarding data analysis. For example, a radiologist can best describe mammography of breast or lung cancer patients. Their guidance helps collect data and annotate the most complex cases. Statisticians contribute valuable insights from the data, while AI engineers build robust and scalable models. Finally, pathologists and practitioners validate the model's outcome, ensuring accuracy and fairness.
However, gathering so many experts is a challenging task. Not only is it time-consuming, but it also adds to the project's cost and complicates the development process.
Lack of transparency
Despite their success in numerous fields, deep learning algorithms are often "black box models”—little information is provided on how the results are achieved.
Transparency and accountability are absolutely crucial for medical experts. Practitioners must provide proper reasoning and justifications for why a certain decision was made. Since AI models do not provide clear reasoning, building trust with their outcomes is difficult.
Digital pathology has significantly impacted medicine and healthcare. Researchers have been developing new methodologies to combat new and existing diseases, adapting techniques like machine learning for improved results and smoother workflows.
The adoption of AI has been mostly well-received throughout the medical research community. Pathologists have started leveraging artificial intelligence for efficiency and improved daily work results. AI models aid scientific research by offering valuable insights and producing excellent diagnostic accuracies from whole slide image analysis.
Moreover, artificial intelligence brings several benefits to pathology, including enhanced cancer diagnosis, offering second opinions for general analysis, training practitioners, and speeding up drug discovery and development.
AI developments in pathology are no short of a revolution in improving healthcare. However, they are not free of challenges. Problems such as a lack of quality data and medical experts for successful training and validation hinder the development of production-grade models. Moreover, the black-box nature of AI raises questions about the model's integrity and creates a trust deficit amongst many users.
Despite the challenges, researchers continue exploring new possibilities and achieving new breakthroughs. The future of AI in digital pathology is looking bright.