Automated wholeslide analysis of multiplex-brightfield IHC images for cancer cells and carcinoma-associated fibroblasts

Use of artificial intelligence in diagnosis of head and neck precancerous and cancerous lesions: A systematic review

“…In six studies, AI-based methods were used to detect oral potentially malignant disorders (OPMD) 52,54,55,57,58,62 with five of these focussing on the detection of oral submucous fibrosis (OSF) specifically. Four studies aimed to detect oral squamous cell carcinoma (OSCC) 53,56,59,60 and one study aimed to classify OPSCC 61 . Overall, seven studies were conducted in India 53,54,55,57,58,59,62 , two in China 60,61 , one in USA 52 and one in Germany 56 .…”

“…Four studies aimed to detect oral squamous cell carcinoma (OSCC) 53,56,59,60 and one study aimed to classify OPSCC 61 . Overall, seven studies were conducted in India 53,54,55,57,58,59,62 , two in China 60,61 , one in USA 52 and one in Germany 56 . Eight studies were published between 2009 and 2015 52,54,55,57,58,60,61,62 and three were published after 2015 53,56,59 .…”

“…These nine methods differ significantly in their learning strategies, as outlined below: ▪ Sugeno Fuzzy 30 A combination of classification and segmentation methods were used to detect potentially malignant and cancerous head and neck lesions (Figure 3). Overall, five studies used AI based classifiers 52,54,56,59,62 , four studies used segmentation methods 53,58,60,61 and two studies used a combination of classification and segmentation methods 55,57 . The most frequently used methods included SVM 36 , Neural Network 38 , Random Forest 32 and clustering.…”

“…The spatial dimensions of images ranged from 262,144 to 10,890,000 pixels; this excludes three studies where the image dimensions were not described 52,56,62 . Although the image sizes are measured in pixels, the actual size of the tissue sample (in microns) will differ due to the varying resolutions of different scanners and the magnification level chosen for the images.…”

“…Baik et al 52 scored the highest for including the use of a separate validation and test set for optimal model selection and evaluation 52 , whereas Sun et al 60 scored the lowest across the 13 areas of assessment, largely due to the description of their approach without the use of reasonable dataset and results 60 . In all applicable studies, except for Lorsakul et al 56 , the methods and intermediate results were clearly presented. Rahman et al 59 was the only study to have used a multi-centric dataset 59 .…”