aetherAI researchers use deep learning to develop a strategy for forming entire slides without annotation

This Article is written as a summay by Marktechpost Staff based on the research paper 'Identification of nodal micrometastasis in colorectal cancer using deep learning on annotation-free whole-slide images'. All Credit For This Research Goes To The Researchers of This Project. Check out the paper, ref blog.

Please Don't Forget To Join Our ML Subreddit

“Deep learning on annotation-free whole-slide images (WSI) has been used to identify nodal micrometastases in colorectal cancer” was published in the peer-reviewed journal Modern Pathology by etherAI, Asia’s leading medical imaging solutions provider specializing in digital pathology and medical imaging AI. With areas under the receiver operating characteristic (AUC) curve of 0.9993 and 0.9956, respectively, the aetherAI algorithm performed well in the detection of macrometastases and micrometastases at the lamina level. For the first time, the companies’ work shows that micrometastases can be recognized using deep learning on whole-slide photos without manual annotation.

The extraordinarily high spatial resolution of whole-slide images makes deep learning of digital pathology (WSI) challenging. Patch-based approaches have been used in most research, which often require full annotation of image patches. On WSIs, this usually involves tedious freehand contouring. aetherAI has developed a method to train neural networks on entire WSIs using only slide-level diagnostics to ease the burden of such an outline and reap the benefits of scaling up training with many WSIs. To circumvent the memory restriction of computer accelerators, they use the unified memory technique.

The extreme resolution of whole slide images (WSI) means dealing with ten billion pixels at the moment, which can easily lead to GPU out of memory issues when training CNN. While traditional patch-based remedies impose the added cost of writing full annotations, aetherAI’s True Gigapixel AI uses full WSIs and slide-level diagnostics to train without annotations.

True Gigapixel AI solves memory limitations and dramatically speeds up pathological AI development by removing contour effort.

This revolutionary strategy can train AI systems using many current WSIs and slides.

aetherAI strives to improve disease detection by using better image recognition capabilities of deep neural networks. Slide quality control, case triage, differential cell counting and IHC quantification are just a few of their services.

Source: https://www.nature.com/articles/s41379-021-00838-2.pdf | Pathologic images of adenocarcinoma and squamous cell carcinoma are shown in a and b, respectively.

Due to the small size of metastatic foci, pathologists find it difficult to detect lymph node micrometastases (tumor size: 0.2 to 2.0 mm). Lymph nodes with micrometastases are counted as positive nodes. Therefore, the search for micrometastases is essential for correct pathological staging of colorectal cancer. Therefore, it would be beneficial to have a pathological staging technique to detect tiny metastatic foci in the lymph nodes. Although deep learning methods have improved the sensitivity and efficiency of detecting micrometastases, manual annotation is time-consuming and labor-intensive. To address this challenge, aetherAI created a deep learning system that identifies colorectal cancer lymph node metastases using its innovative end-to-end training technique using annotation-free WSIs.

Chang Gung Memorial Hospital in Taoyuan, Taiwan, collaborated on the research with pathologists. There were 1963, 219, and 1000 slides in the training, validation, and test sets.

Experiments on a WSI dataset of lung cancer show that the suggested technique obtains areas under the receiver operating characteristic curve for classification of adenocarcinoma and squamous cell carcinoma on 0.9594 and 0.9414, respectively. . Moreover, by using class activation mapping, the approach achieves better classification results than multiple instance learning and strong localization results for minor lesions.

Conclusion:

The deep learning model developed by aetherAI uses whole-slide images of regional colorectal cancer lymph nodes with only a slide-level label (positive or negative slide). A deep learning model was trained to detect metastases using a TAIWANIA 2 supercomputer. The system worked well to see macrometastases and micrometastases at a single lymph node, with ASCs of 0.9993 and 0.9956, respectively. The aetherAI model correctly recognized nodal metastases based on tumor cell regions, validated by visualization using class activation mapping.