Researchers developed NAS-SGAN, an AI solution for breast cancer grading, achieving 98% accuracy using only 20% labeled data. Traditional GANs detected cancer but couldn’t classify it; NAS-SGAN does both by combining labeled and unlabeled images. The model couldn’t run on GPUs due to memory limits, so Intel supported its implementation on 2nd Gen Intel® Xeon® Scalable processors with 192GB memory per server. Using Intel® Optimization for TensorFlow and Horovod for distributed training, researchers trained the model efficiently on high-resolution images. This approach improves grading consistency, streamlines diagnosis, and reduces annotation workload.

Join for free to read