Weakly-supervised tumor purity prediction from frozen H&E stained slides

Weakly-supervised deep learning models in computational pathology

“…In this weakly-supervised approach, Brendel et al, employed an attention based, multi-task, multiple-instance learning (MIL) model to learn weight features for ROIs within a slide as well as feature representation that can vie with pathologist-derived estimates of tumour purity, exceeding accuracy of previous supervised learning approaches. 4 With tumour purity associated with tumour type, the author's model could predict cancer type in both test and validation tests with 93% accuracy. However, misclassification of breast cancer and lung cancer were common.…”

“…This is regarded as a more time-consuming, and computationally expensive approach that is dependent on large, gigapixel sized images or extensive pixel-level annotations. 2 , 4 , 5 To overcome these limitations, research is being undertaken in developing weakly-supervised approaches to deep learning where the slide is given a single annotation (label) with features from image patches or tiles being pooled under a multiple-instance learning framework (MIL). Thus with slide-level labelling if a slide is positive then one or all tiles must contain a tumour sample, whereas if a slide is negative all tiles must be tumour-free.…”

“…Thus with slide-level labelling if a slide is positive then one or all tiles must contain a tumour sample, whereas if a slide is negative all tiles must be tumour-free. 4 , 5 …”

“…A recent paper published in eBioMedicine, Brendel et al ., 4 using data accessible from The Cancer Genome Atlas (TCGA), presents a weakly supervised deep learning model that is capable of identifying key features on H&E stained slides to accurately estimate tumour purity. Tumour purity refers to the percentage of tumour cells over a range of regions of interest (ROI) in a tissue section enumerated by pathologists.…”

“…However, misclassification of breast cancer and lung cancer were common. 4 While this may relate to the spatial distribution of the tissue, reducing noise and variability in generation of large datasets and ensuring robusticity of algorithms, is crucial to preventing false negatives and positives in real-life application. 4 , 8 The model derived by Brendel et al additionally was able to permit visualisation of ROIs.…”