At the forefront of AI research in medicine are diagnostics and drug development. Deep learning technology allows better decision making, improved efficiency in clinical trials, and a clearer path to drug development. Keen Eye applies deep learning technology to imitate and enhance the visual expertise of a medical expert. Compared to the traditional image analysis approach, deep learning requires minimal input from pathologists. In a recent study, we applied deep learning technology to subtype lung cancer samples with excellent results. This AI model, that can reliably diagnose the cancer type, can also be applied to drug discovery in identifying new patterns associated with clinical outcomes, hence accelerating drug development.
Study of deep learning to subtype lung cancer cases
Using non-Small Lung Carcinoma (NSLC) cases from the public data set, TCGA, we trained a deep learning model to classify the lung cancer subtype, squamous carcinoma and adenocarcinoma, of 316 cases. The model predicted the clinical outcome with an accuracy of 0.883, which is comparable to the accuracy of a trained pathologist.
The model also generates a heatmap highlighting regions of interest that it used to classify the slide. We randomly selected 30 correctly classified WSI – 15 from each type of cancer. Using the heatmap with the hotspots as a guidance, a pathologist reviewed each slide. He qualitatively assessed the cases and confirmed the model’s ability to correctly identify the relevant region that characterizes the specific lung cancer subtype.
This study demonstrates the use of deep learning AI to accurately diagnose and subtype lung cancer cases using whole slide images with only one annotation per patient.
Benefits of deep learning
Deep learning offers unique innovations in augmenting human expertise in biomarker quantification and discovery, whether in the clinical setting or in drug development research. The deep learning model delivers these concrete benefits:
The deep learning model requires minimum human input
Our study demonstrated that using a single, slide-level annotation, i.e. the diagnosis confirmed with clinical outcome, the deep learning model is able to classify between adenocarcinoma and squamous carcinoma subtypes in lung cancer. It does not require a pathologist to laboriously annotate differentiating tissue types or lesions, nor an image analysis programmer to develop feature-based algorithms. This makes the deep learning AI model more efficient and accessible to pharmaceutical companies.
The deep learning model can learn using limited data
In the case of lung cancer subtyping, when a pathologist is unable to render a diagnosis based on the histopathologic sample, he would often request additional samples such as clinical or genomic data to deliver a final analysis. In our study, the deep learning model demonstrated the ability to deliver an accurate classification using information from the tissue sample alone.
The deep learning model is a valuable assistant
The model uses an intelligent attention mechanism to assist the human expert. It highlights the regions of interest (ROI) that contributed most to the final decision. This can either help the pathologist to confirm the classification quickly or in the case doubt, to investigate further. This intelligent AI Hotspot removes the ‘black box’ mystery that is often associated with deep learning models.
How deep learning can be applied to drug discovery
The same way deep learning technology can distinguish between squamous cell carcinoma and adenocarcinoma, it holds great potential to find differences in biomarkers and histological features between
- samples that are responsive versus refractory to therapy
- samples before and after treatment
- healthy and pathological samples
- different disease stages
Digital pathology’s potential for improving diagnosis and drug discovery and development, using AI, is poised to greatly improve patient care. Deep learning is rapidly advancing many areas of science and technology, and Keen Eye has the tools and experience to partner with pharma companies and makes an impact in the diagnostic and pharmaceutical industries; particularly in the fight against cancer.