Predicting Mutated Genes in Multiple Myeloma from Blood-Smear Whole-Slide Images- An Early Exploration Model
Afya Shaikh
Lay Summary:
I built a machine-learning model that scans microscope images of blood and flags spots that might contain myeloma, a type of blood cancer. It’s an early assistive tool to help pathologists find concerning areas faster, and with more training it could speed up detection and care, especially where lab resources are limited.
Abstract:
Identification of multiple myeloma on blood-smear or bone-marrow whole-slide images (WSIs) is time-consuming and subjective in early diagnosis. This study will assess whether a small convolutional segmentation technique is capable of predicting myeloma-rich areas on WSIs. WSIs were tiled into 512×512-pixel patches and split 80/20 into training and validation sets (batch size 2). Light U-Net–variant convolutional neural network performed binary segmentation (background/myeloma) and was optimized with Adam (learning rate 5×10⁻⁴) using composite loss of class-weighted cross-entropy and Tversky (α=0.8, β=0.2). Accuracy was evaluated by precision, recall, F1 score, accuracy, area under the receiver operating characteristic curve (AUC ROC), and confusion matrix. Accuracy in validation data was 0.79, and the AUC was 0.92. Whole results were- background—precision 1.00, recall 0.78, F1 0.88; myeloma—precision 0.18, recall 0.98, F1 0.30, which shows high sensitivity but many false positives under extreme pixel-level imbalance. This model is useful to pathologists because it can indicate suspicious regions and help reduce time and errors. It can be improved with rebalancing and enhancing the data, tweaking thresholds and training parameters, trying stronger encoder backbones, training longer, and validating on a patient level so that it can generalize to new cases.
Q&A:
Bios: Afya Shaikh
Program Track: Advanced Research
GitHub Username:
AfyaS -Afya Shaikh
What was your favorite seminar? Why?
My favorite seminar was Marietta Montivero’s because she’s doing exactly what I hope to pursue in the future. She combined dermatology, skin cancer research, and machine learning all together. It was inspiring to see how passionate she is about her work, and it made me even more excited about the possibilities ahead in my own career. -Afya Shaikh
If you were to summarize your summer internship experience in one sentence, what would it be?
It was a valuable and enjoyable learning experience where I gained new coding skills and felt well-supported throughout. -Afya Shaikh