Capricorn AI: An Automated Deep Learning Approach for Histopathological Tissue Classification

Moshe Newman, Verrazzano Class of 2025, completed major in Molecular & Cellular Biology

I identified my research topic at the intersection of oncology, bioinformatics, and artificial intelligence, motivated by my longstanding passion for cancer research and precision medicine. My goal was to contribute toward improving the diagnostic accuracy and efficiency of cancer detection, ultimately aiming to benefit patient outcomes. The idea of utilizing advanced technology like deep learning to tackle histopathological classification inspired me, especially since the method holds potential for significant clinical impact.

Early on, I had expected the capstone to be straightforward training and testing of deep learning models. Instead, it turned out to be far more complicated and involved frequent troubleshooting and optimization. The capstone required heavy preprocessing, model architecture exploration, and close attention to model outputs. The complexity of converting results into clinical understanding was deeper than anticipated but ultimately more rewarding.

Among the greatest challenges was handling dataset imbalances and hyperparameter tuning of the neural network to avoid overfitting, and it took a lot of experimenting and statistical exploration. In contrast, understanding the theoretical background of deep learning was relatively easier to me given my background in bioinformatics as well as programming. What was most surprising to me was the complexity involved in adequately visualizing and representing the model output predictions, which necessitated creativity and more statistical expertise than expected.

To further expand this work, I plan to incorporate patient metadata and clinical history to enhance the predictive capability of Capricorn AI, effectively making it a more detailed diagnostic and prognostic tool. Exploring real-time imaging and adaptive training techniques could significantly improve clinical utility and specificity of the model. Lastly, conducting larger validation studies with more varied datasets will be necessary to facilitate generalizability.

Through this experience, I am developed a greater appreciation of the challenges and opportunities of interdisciplinary research. It reinforced my passion for bioinformatics and oncology and expanded my expertise in machine learning, particularly the importance of meticulous data handling, model verification, and successive experimentation. Professionally, it established my analytical skills, research endurance, and ability to present sophisticated scientific outcomes succinctly and persuasively to different audiences.