Deep Learning Algorithms for Detection of Lymph Node Metastases From Breast Cancer
AI progress in healthcare is slower than other domains but pace of implementation is as high as 600M$ in 2014. Expected to grow up to 6.6B$ in 2021. Most enhancements in the latter domain is in the Image Analysis field focusing on feature engineering into Deep Learning Models. In a study, authors compared CNN’s ability to discriminate between different types of skin cancers including malignant melanoma. They were almost on par with 21 board certified dermatologists in evaluating. Authors suggested that models could be deployed on smaller devices to give low cost access around the globe. In another study, a Deep CNN was applied on a test set of more than 128000 adults’ retinal fundus with diabetes to identify referable diabetic retinopathy and had a very high sensitivity and specificity for detecting referable diabetic retinopathy and macular edema. Also made clear that AI was not to replace physicians but to enable simple and high level testing at a faster pace. Since radiology was already in position with image analysis, it made AI deployment easier to port into different types of scans like breast lesions, pulmonary nodules etc. Radiology has always been off late to implement digital imaging and thus, computer-assisted diagnostic tech. Conversion to digital images eliminated film, chemicals, developers and film storage. Their management was also not very easy so its elimination increased a lot of potential in usage of computers.
Digital pathology has realised many practical benefits but also needs a lot of workflows especially on a very data heavy ecosystem. In April 2017, Philips received US FDA clearance for a product used for reviewing scanned glass pathology on a monitor. AI in healthcare is being implemented on a pre-built base of digital datasets and usable digital images. In a major contest called CAMELYON16, A Deep CNN performed better than other algorithms. For a study, 2 approaches were taken to study the analysing capacity of a group of variably experienced pathologists. First method was to analyse and review close to 129 slides for which the group had to finish each slide in under a minute and the other method was to give the same set of slides to a pathologist with no time limit and undoubtedly, the latter gave better results. The metastases were the main feature that defined the confidence in algorithms. Although metastases are still being studied on, the fact that algorithms did undoubtedly better than pathologists is exciting. However, the limitations end at detecting lobular carcinoma, for which, the pathologists are still unparalleled at. Yu et al’s fully automated informatics pipeline to extract objective quantitative image features and build classifiers to distinguish lung cancer with different survival outcomes was able to predict better outcome than either clinical stage or histopathological grade. It also achieved similar results with squamous cell carcinoma of the lung.
A major unresolved issue is how AI will be implemented in routine clinical practice because it will have to address several significant obstacles. The first would be to create value proposition in pathology and the other is cost. AI must be able to demonstrate flexibility, efficiency and also safety. The cost factor is something of a function of value proposition if its value is demonstrated and it results in government and third-party encouraging the use of AI in pathology. Though there are organisations and activities that support it, it is usually looked down and rejected. Third would be the education factor which requires a lot of AI education and other computational methods to be taught to pathologists in training and also requires them to be comfortable to use digital images and other data and features in combination with algorithms in daily life and to be optimistic, it would take around 5-10 years to implement and build such a workforce.
AI as an idea has become a major element of the healthcare landscape, and AI as a reality can potentially provide value in many sectors. While still requiring evaluation within a normal surgical pathology workflow, deep learning has the opportunity to assist pathologists by improving the efficiency of their work, standardizing quality, and providing better diagnosis. Though such advancements might reach human level, there is a very little risk of pathologists being replaced as their cruciality will still remain though only their new workflow is going to be the future.
Comments
Post a Comment