Lymph Node Metastases Detector

About

High-quality pathology reporting, which is heavily dependent on histopathologists’ expertise, is a central part of the diagnostic pipeline and critical for accurate and timely diagnosis and patient’s management. Unfortunately, the pathology workforce is insufficient to meet the diagnostic demand. Pathology requests increase at a steady rate of 4.5% per year whilst the pathology workforce increases by only 1.2%. In the UK 97% of pathology departments have staff shortages. An estimated annual £27m is spent by the NHS to cover this gap through overtime, Waiting List Initiatives, outsourcing and the use of locums. These solutions result overall in lower quality diagnoses, higher misdiagnosis rate, diagnostic delays, worse patient outcomes and lower patient quality of life. Moreover, misdiagnosis-related costs for the NHS including unnecessary or wrong treatments, prolonged hospitalisation periods, other various inefficiencies and litigation costs was estimated at £128.9m in 2017. DeepMed IO utilised NHS funding in two phases through the Small Business Research Initiative to develop DeepPathTM LYDIA, a decision support system for detecting metastasis on Hematoxylin and Eosin (H&E) Lymph Nodes (LN) from oncology patients, based on Convolutional Neural Networks trained under the supervised learning paradigm. We decided to pursue a system that would increase the efficiency in the above task for 3 reasons: (a)Diagnosis of metastases on LNs is performed for all solid tumors (b)It directly affects the determination of disease staging which in turn defines (along with grading) patient management and the selection of therapeutic strategy (c)It is a labor-intensive process that consumes a significant amount of pathologists’ time. A pathology lab in the UK produces on average 150,000 slides p.a, 20% of which contain LNs. In the UK there are ~150 pathology labs generating 4.5m LN slides pa in total. Given that a pathologist needs 3 minutes on average per slide, the examination of all the LN slides costs the NHS 1,407 Person Months pa. LYDIA comprises of a case management/viewer module and AI modules for analysing the digitised microscope slides (Whole Slide Images - WSI) and performing the detection of metastatic lesions. These lesions are presented to the pathologist as outlined segments along with specific measurements, such as the maximum segment diameter and the tumor probability. Furthermore, at the case level all the LN-WSIs that belong to a single case are re-ranked according to the detected tumor content of each WSI showing the ones with the highest tumor content first. Since LYDIA is not an academic project but one leading to a commercial product, it is optimised for data-security and speed. It is capable of analysing a single WSI in under 5 minutes on a local GPU-server with a minimum of two GPUs. Furthermore, the system is web-browser based, hence platform agnostic, it can be as easily deployed through the cloud with load-balancing capabilities for seamless and scalable operation and can process images from all the main WSI-scanner vendors. Additionally, we have secured technical partnerships with vendors of digital pathology Picture Archiving and Communication Systems (PACS) such as Inspirata, for delivering our technology through their platforms, thus ensuring seamless integration of LYDIA in the digitised pathology lab work-flow. Finally, DeepMed IO has been selected as an AI solution-provider for the Innovate UK funded PathLake+ project (https://www.pathlake.org/), which has undertaken to fully digitise the pathology labs of more than 25 hospitals in Southern England (excluding London), where LYDIA modules (different tumor types) will be installed and tested in real-life data. LYDIA along with the first AI module for detecting metastasis detection on LNs from breast cancer patients are at the final stages of fine-tuning and testing leading to a CE-Mark, expected towards the end of Q1-2021. The model for breast cancer is 97.25% sensitive at the LN-level and detects an average of 12 False Positive Segments per Lymph node. The above metrics have been derived by analysis for 182 positive LNs from 4 different centers. We also performed a study for quantifying the efficiency-gains when using the system as decision support. As such 70 LNs (20 macro metastasis , 20 micro metastasis, 10 isolated tumour cells (ITC) and 20 negative) were examined by 20 histopathologists with and without AI-augmentation under a relaxed time constraint of 3 minutes per WSI. AI-decision-support delivered an average of 2.2 times increase in diagnostic speed, broken-down as follows: 2.28 times for macro-WISs, 4.98 times for micro-WSIs, 2.38 times for ITC-WSIs and 1.38 times for negative-WSIs; there was also a 4% increase in diagnostic accuracy as per the Cohen’s correlation coefficient when the AI-decision support was used. On a different note, we developed a secondary AI-model to determine LN-level labels indicative of whether macro, micro-metastasis or ITCs were detected anywhere on the WSIs. This combined pipeline achieved a Cohen’s kappa-correlation of 90.52% in the automatic diagnosis of the nodal stage in the international competition Camelyon-17. At the time of submission DeepMed IO achieved the 5th best-performance globally out of 120 submissions made from 62 teams. However, in order for a metastasis detection system to maximise its utility and hence be adopted by any health-care system it ought to be pan-cancer. The reason is that by limiting the model to a specific tumor type, its utility is like-wise limited to the prevalence of the specific tumor type and furthermore, the sensitivity of the model even for a particular tumor type cannot be guaranteed (a) because there might be extremely rare tumor sub-types that have not been adequately represented in the training set and (b) the tumor site under-investigation might not be the primary one; hence a tumor that was located in the breast might be a lung-tumor that has metastasised. Bringing however a pan-cancer metastasis detection system to market under the supervised-learning paradigm is a particularly challenging task since thousands of slides for all the different tumor types should be collected, quality controlled and annotated for training as many different models. As such and although AI-modules for other cancer types such as colon and lung cancers as well as melanoma are already under development, we have focused our main efforts in utilising convolutional autoencoders through the anomaly-detection paradigm in order to deliver the first pan-cancer metastasis detection model in the world. Through extensively training a model to recognise the normal LN tissue architecture, we will deliver a sensitive detector of any tumor type, no matter how rare the subtype is, since the detection is based on the deviation from learned normality. Up to now we had extremely positive results achieving statistically significant results with very high confidence in the detection of metastases from breast, colon, lung, melanoma and gastroinstestinal cancers. We project to have completed the model training & evaluation and file for CE-Marking in Q3-2021.