Brain Tumor MRI Deep Learning Detection & Segmentation - CSCI E-25 Computer Vision - HarvardU
[Brain Tumor Deep Learning Detection (CNN, VGG, ResNet, EfficientNet, ConvNeXt) & Segmentation (U-Net, DeepLabV3+, K-Means, Segment Anything Model - SAM)]
I’m excited to share a short YouTube presentation of one of the most ambitious and technically complex projects I’ve worked on so far—my Computer Vision final project for CSCI E-25 by Professor. Stephen F. Elston, PhD, focused on deep learning–based Brain MRI tumor analysis using a 2-stage pipeline: classification first, followed by semantic segmentation.
Earlier this year, during the Spring term (January–May), I decided to challenge myself by turning that motivation into action. This project was intentionally designed as a full end-to-end system rather than a single model. I first built binary and multi-architecture classification pipelines to determine the presence of a tumor, using transfer learning and fine-tuning across several modern CNN families. Only after achieving reliable classification performance did I move to pixel-level semantic segmentation, where the focus shifted from detection to precise tumor localization and boundary delineation.
This staged design added both realism and complexity. It required careful coordination across preprocessing, model assumptions, data consistency, and evaluation strategies, while avoiding leakage between stages. I implemented and compared a diverse set of architectures, including ResNet, EfficientNet, and ConvNeXt for classification, and U-Net and DeepLabV3+ for segmentation. Performance was evaluated using a combination of accuracy, ROC-AUC, IoU, confusion matrices, and qualitative visual inspection of segmentation outputs to fully understand model behavior.
I’m deeply grateful to Professor. Stephen F. Elston, PhD, for his guidance and support throughout the Spring term. His advice on leveraging pretrained models, structuring experiments, and handling medical and grayscale imaging challenges played a critical role in helping me navigate the complexity of this work and achieve strong, stable results.
This project reinforced several important lessons for me. Complex systems benefit from staged design rather than monolithic models. Strong classification is a prerequisite for meaningful segmentation. In medical imaging, preprocessing rigor and evaluation discipline are just as critical as architecture choice.
Your feedback is very welcome. Thank you.