Most cancer specialists and radiologists rely mainly on histopathological reports to detect early symptoms of breast cancer. According to the National Cancer Institute, it has been estimated that around 20% of cancer signs are missed when depending only on these results and human skills. This is where Al plays an important role. As such, more hospitals are choosing the path of AI software to assist doctors in their work. In fact, Intel and Penn Medicine conducted a study that revealed the emergence of Al in the medical field helping doctors increase the detection of brain tumors by 33%.
Dr. Madhu Nair and Dr. Asha Das, two scientists from the Artificial Intelligence and Computer Vision Lab at Cochin University in India, have created a technique that can identify breast cancer at its early stages. They took the help of Intel-based software and hardware to successfully complete the project. Known as the NAS-SGAN model, it is said to detect cancer based on certain labeled and unlabeled medical images. The model utilizes deep learning to differentiate between the various cancer grades.
Tackling the problem
Nair and Das initially faced a problem because of the GPU’s inability to store the entire Al model in memory. This led to some issues in the interpretation of high-resolution photos that were required as part of the deep-learning solution. During mid-2022, Nair was scheduled to attend a meeting with a Dell representative who recommended the Intel India team to help complete the project. Analyzing the issues, Intel decided to offer four servers with its Xeon Scalable processors, with each of them having a storage memory of 192 GB. These servers run as a single compute cluster and are linked to the storage using a high-speed Ethernet network.
The NAS-SGAN model was also designed to operate in the absence of any deep learning accelerators. As for the software, it uses Intel Optimisation for TensorFlow which makes it easy to manage acceleration features in Intel CPUs. Nair praised the Intel team by saying that they were the only ones who immediately recognized the significance of the project.
Optimistic results
The results obtained through the NAS-SGAN model were beyond the expectations of the researchers, with an additional bonus. Apart from detecting signs of cancer, the technology was also capable of differentiating various stages of cancer. It achieved an accuracy rate of 98%, approximately 10% higher when compared to other similar models.
The results seem to be more impressive because the model was used only on a limited amount of annotated data. It is expected that the accuracy rate can rise even higher when the model is tested on more datasets. Nair and Das are now planning to develop a similar technique for cerebral aneurysms and identifying polyps from endoscopies.
Photo credits: The images used are owned by Intel and have been provided for press usage.
