Fully automated volumetric measurement of malignant pleural mesothelioma by deep learning AI: validation and comparison with modified RECIST response criteria

Andrew C Kidd ¹, Owen Anderson ^2 3, Gordon W Cowell ⁴, Alexander J Weir ³, Jeremy P Voisey ³, Matthew Evison ⁵, Selina Tsim ^1 6, Keith A Goatman ³, Kevin G Blyth ^{1 6 7}

Measuring treatment response in mesothelioma is hard. The complex shape of the tumour makes surrogates of true volume (eg mRECIST) unreliable but measuring volume is too laborious for humans. We’ve just published the 1st automated mesothelioma AI in Thorax doi: 10.1136/thoraxjnl-2021-217808

This project was made possible by funding from @cancerchallscot @CSO_Scotland @BlfResearch, support from @DataLabScotland and a brilliant collaboration with @CanonMedicalEDI. We’re continuing that partnership in @PREDICT_Meso, working with @NCIMImaging on large-scale validation.

The study required generation of precise human ‘ground truth’ (reference tumour annotations to train the AI). That meant drawing round tumour on every slice of every CT (ave 225 slices/scan, 140 scans fully annotated). A huge effort by Andrew Kidd.

Creation of the AI was led by Keith Goatman @CanonMedicalEDI and the fantastic Owen Anderson. Great support form @stucdubh @jpvoisey and fantastic clinical team @SelinaTsim @MatthewEvison1 – especially radiologist @GWCowell

1. Glasgow Pleural Disease Unit, Queen Elizabeth University Hospital, Glasgow, UK
2. School of Computing Science, University of Glasgow, Glasgow, UK
3. Canon Medical Research Europe Ltd, Edinburgh, UK
4. Department of Imaging, Queen Elizabeth University Hospital, Glasgow, UK
5. Department of Respiratory Medicine, University Hospital of South Manchester, Manchester, UK
6. Institute of Cancer Sciences, University of Glasgow, Glasgow, UK
7. Beatson Institute for Cancer Research, Glasgow, UK