FB Pixel no scriptAlibaba’s Damo Academy introduces AI model that can diagnose eight cancer types in a single analysis | KrASIA

Alibaba’s Damo Academy introduces AI model that can diagnose eight cancer types in a single analysis

Written by KrASIA Connection Published on   3 mins read

The new model is currently undergoing testing in China, but holds the potential to enhance the precision of cancer diagnostics.

Alibaba’s Damo Academy, the science and innovative technology research institute of Alibaba Group, has introduced an artificial intelligence model that can help doctors to identify, distinguish between, and diagnose eight of the most common and deadliest types of cancer. This development is expected to improve the accuracy of cancer diagnostics and reduce the risk of erroneous results.

Medical AI models possess great potential in recognizing organ-specific diseases and assisting doctors in making accurate diagnoses. However, significant challenges arise when they attempt to identify multiple health issues across multiple organs.

Minimizing the risk of erroneous results is one example of the challenges involved. In particular, false positives—identifying that an issue is present when it is not the case in reality—and false negatives as well as missed diagnoses can be critical for patients, especially those with multiple cancers. To minimize this risk, doctors often have to conduct comprehensive examinations, underscoring the pressing need to develop a more accurate and efficient model of diagnosing multiple types of cancer.

The Damo Academy, in collaboration with other medical institutions such as the Sun Yat-sen University Cancer Center, has attempted to fulfill this need by developing a multi-cancer image analysis model called cancerUniT.

The model is built on Mask Transformer, a semantic segmentation technique that annotates image data at pixel level. Mask Transformer is an example of AI architecture designed to analyze images at a granular (pixel) level to facilitate algorithmic identification of the elements portrayed.

In cancerUniT’s case, the model is used to analyze images of body tissues for eight cancer types, including lung, colorectal, liver, stomach, breast, esophagus, pancreas, and kidney cancer. It is also able to recognize specific subtypes of tumors within infected organs.

Photo of the model identifying eight different types of cancer. Photo source: Alibaba Damo Academy.

The complexity of multi-cancer diagnostics lies in the interconnectedness of various organs, malignant tumors, and other tumor types. For example, liver cancer and liver cysts reside in the same organ, yet they may exhibit disparities in texture and malignancy characteristics. Similarly, liver cancer and pancreatic cancer resemble each other in form and structure but are distinct malignant cancers situated in different organs.

The Damo Academy used the Transformer model, a neural network that can learn context and meaning by tracking relationships in sequential data. To facilitate a more effective learning process, the institute developed a method for the model to learn about the unique traits and patterns of each tumor type. The model is also trained to consider the relationships between different types of cancer and their subtypes within various organs, improving the consistency and reliability of diagnostics.

In a comparative study involving 631 patients, the new model outperformed traditional models. The new model displayed higher sensitivity in detecting and segmenting tumors, at 93%. It also displayed a good level of specificity, correctly identifying non-cancerous areas 82% of the time.

According to Lv Le, head of Alibaba’s Damo Academy’s medical AI team and fellow at IEEE, the new model marks a significant milestone in medical technology. Being able to diagnose the eight most deadly types of cancer in a single analysis not only simplifies the diagnostic process, but also improves its accuracy. In particular, this technology will benefit radiologists, who will be able to use it to identify cancers that have recurred or spread to other parts of the body.

The research paper detailing this new model’s capabilities has been recognized by the 2023 International Conference on Computer Vision, and the technology is currently being tested in several hospitals including Shanghai General Hospital.


Auto loading next article...