E0476 - Comparison of Radiomic Machine Learning and Deep Learning Survival Models in Patients With Hepatocellular Carcinoma

Tuesday, October 25, 2022

3:00 PM – 5:00 PM ET

Location: Crown Ballroom

Has Audio

Presenting Author(s)

Shivabalan Kathavarayan Ramu, MBBS

Mayo Clinic
Rochester, MN

Shivabalan Kathavarayan Ramu, MBBS¹, Anjali Byale, MBBS¹, Achintya Singh, MBBS²
¹Mayo Clinic, Rochester, MN; ²Cleveland Clinic, Rochester, MN

Introduction: To compare machine learning (ML) and deep learning (DL) approaches to predict survival outcome in patients with Hepatocellular Carcinoma (HCC).

Methods: The clinical and image data of 82 patients were accessed from the TCGA-LIHC collection in The Cancer Imaging Archive. Patients with incomplete image sequences were excluded to finally include 39 patients. We delineated the following regions of interest: 1) Peritumoral 1 (first 2mm zone around the tumour), 2) Peritumoral 2 (second 2mm zone around the tumour), 3) Tumoral, 4) paravertebral muscles L4, L5, 5) Paravertebral muscles L1, L2, L3, 6) Psoas muscle at the L1, L2 & L3 levels 7) rest of the psoas muscle. The radiomics features were extracted from the CT images using the LIFEx software(www.lifexsoft.org). We evaluated the following models: 1) Cox proportional hazards model by componentwise likelihood based boosting with stepnumber 10 and penalty number 100, 2) Deephit: DeepHit: trains a neural network to learn the estimated joint distribution of survival time and event, while capturing the right-censored nature inherent in data. Analysis was done with frac 0.3, relu activation, 0.1 dropout, 100L epochs and a batch size of 32L, 4) Multitask logistic regression model with ranking based feature selection to predict survival using a logistic regression model and the parameters from each model are estimated simultaneously in the maximization of the joint likelihood function, and 5) random survival forest with 1000 trees. Analyses were done in RStudio, and missing values were imputed using missRanger package.¹ The data is split into 80 percent training and 20 precent validation. All models were 5-fold cross validated. Prediciton statistics was calculated for each model developed.

Results: Out of the selected ML and DL models, Multitask logistic regression preformed the best in classifying all regions of interest. Amongst the regions of interest, paravertebral muscles and psoas muscles were trained with higher AUCs than the actual tumour regions themselves.

Discussion: CT-derived radiomics were valuable for noninvasively assessing the survival and also the paravertebral and psoas muscles have better predictive capacity to predict survivals apart from the tumour or peritumoral regions themselves.

S.No.	Model specification	AUC	C-i ndex	AUC	C-i ndex	AUC	C-i ndex	AUC	C-i ndex	AUC	C-i ndex	AUC	C-i ndex	AUC	C-i ndex
Region of interest		1	1	2	2	3	3	4	4	5	5	6	6	7	7
1.	Cox boost	0.47	0.45	0.60	0.71	0.73	0.79	0.46	0.63	0.40	0.46	0.60	0.59	0.66	0.71
2.	Deep hit	0.47	0.45	0.60	0.71	0.73	0.79	0.46	0.63	0.40	0.46	0.60	0.59	0.66	0.71
3.	Multitask logistic regression	0.87	0.81	0.86	0.93	0.80	0.80	0.93	0.89	0.93	0.89	0.86	0.93	0.86	0.93
4.	Random survival forest	0.5	0.5	0.5	0.5	0.5	0.5	0.5	0.5	0.5	0.5	0.5	0.5	0.5	0.5

Table: Table 1. Performance of machine learning and deep learning models for radiomic survival prediction

Disclosures:

Shivabalan Kathavarayan Ramu indicated no relevant financial relationships.

Anjali Byale indicated no relevant financial relationships.

Achintya Singh indicated no relevant financial relationships.

Shivabalan Kathavarayan Ramu, MBBS¹, Anjali Byale, MBBS¹, Achintya Singh, MBBS². E0476 - Comparison of Radiomic Machine Learning and Deep Learning Survival Models in Patients With Hepatocellular Carcinoma, ACG 2022 Annual Scientific Meeting Abstracts. Charlotte, NC: American College of Gastroenterology.