Table 1.

Summary of studies on computer-based methods included in the review.

Publication	Computer Based Method	Types of Data	N Patients	Problem/Assignment	Validation Method	Accuracy/Results
Chaisaowong K. et al., 2007 [18]	A convexity model with a Hounsfield Unit threshold was implemented to detect pleural thickness	CT	3	To develop an automatic image processing approach to detect and quantitatively assess pleural thickenings
Pena E. et al., 2017 [19]	Combinations of radiomics features used to generate logistic regression models. 3 texture and 3 shape features	CT and MRI	34	To identify a radiomic approach that may help differentiate benign versus malignant pleural lesions	Visual assessment by thoracic and abdominal radiologists	CT model revealed an AUC of 0.92 ± 0.05 outperforming abdominal radiologists
Pavic M. et al., 2018 [20]	An in-house developed software was implemented to extract features from manually contoured CT scans	CT	11	To investigate the impact of inter-observer variability in manual tumor delineation on the reliability of radiomic features	3 experienced radiation oncologists manually segmented CT scans	Median Dice Sørensen coefficient (DSC) was low (0.26) with a low stability rate of radiomic features (36% of total parameters)
Gill R. R. et al., 2012 [21]	Selective segmentation with the 3D feature of the software, with manual segmentation of extrapleural sites of disease	CT	338	To assess the usefulness of CT-derived tumor volume for stratifying survival after surgery-based multimodality treatment		At multivariate analysis a tumor volume > 500 cm3 showed a HR = 2.02, p = 0.0109
Pavic M. et al., 2020 [22]	Extraction of CT and FDG PET features to build a Cox regression model	FDG PET and CT	123	To build a CT and FDG PET radiomics model for the prediction of prediction free survival (PFS) in MPM		Concordance index of 0.66 was obtained for the PET radiomics model, CT radiomics model not successfully validated
Labby Z. E. et al., 2013 [23]	Semiautomated segmentation with semiautomated shape-based interpolation requiring seeding	CT	81	To create a comprehensive model for MPM survival utilizing continuous, time-varying estimates of disease volume from CT imaging in conjunction with clinical covariates		Final multivariate survival model included continuous specific growth rate from baseline (HR = 1.31)
Fan Liu, et al., 2010 [24]	Semiautomated segmentation method combining chest-rib interpolation, gradient vector flow snake and multiple thresholding technique with manual editing of suboptimal results by a thoracic radiologist	CT	30	To calculate the tumor volume and to investigate whether the baseline volume or volume change after chemotherapy predicts patient survival	A second radiologist independently reviewed the computer results	Percentage change of tumor volume from baseline to first follow-up CT was significantly associated with overall survival (HR = 1.94)
Labby Z. E. et al., 2013 [12]	5 observers manually contoured tumor on 3 selected sections, then contours were converted to area measurements using Green’s theorem	CT	31	To evaluate manual area measurements as an alternate response assessment metric, specifically through the study of measurement interobserver variability		The time required to contour tumor for each scan was 20 min; the 95% CI for relative interobserver variability for summed area measurements was [−71%, +240%] for baseline and [−41%, +70%] for FU scan
Frauenfelder T. et al., 2011 [25]	A dedicated software semiautomatic feature with linear interpolation was implemented to segmentate MPM and compute the tumor volume	CT	30	To assess robustness of volumetric measurement of MPM before and after chemotherapy compared to mRECIST criteria	3 readers independently assessed the tumors response	For tumor volume compared to mRECIST were found a high inter-rater reliability (0.99) and inter-observer agreement (general k 0.9)
Armato III S. G. et al., 2004 [26]	6 computerized algorithms (from Minimum-distance algorithm to Normal-to-initial-end-point algorithm) given a specified initial endpoint measured tumor thickness	CT	22	To evaluate the variability of manual MPM thickness measurements in CT scans and to assess the relative performance of six computerized measurement algorithms	5 observers manually measured tumor thickness	Computer based tumor thickness measurements highly correlated with the average of observer measurements (R ≥ 0.93)
Armato III S. G. et al., 2005 [27]	A semiautomated method computes tumor thickness requiring the manual selection of a point in the outer margin of tumor	CT	22	To evaluate the clinical acceptability of semiautomated methods for the measurement of MPM thickness in CT scans	3 radiologists and oncologists independently reviewed measurements	86% of semiautomated measurements were accepted without modification
Sensakovic W.F. et al., 2011 [28]	An automated method based on grey level, texture and shape analysis segmented lung and nonlinear diffusion and a k-means classifier identified MPM in the pleural space	CT	31	To present a computerized method for the three-dimensional segmentation and volumetric analysis of MPM	3 observers independently contoured 5 randomly selected sections for each scan	The median Jaccard index between the computer based and manual segmentation was 0.484
Chen M. et al., 2017 [29]	A random walk-based algorithm was implemented to segment the tumor	CT	15	To assess the performance of a computer-aided semi-automated algorithm for the purpose of segmenting MPM on CT	Manual delineation by a clinical radiologist	A mean DSC of 0.825 was achieved; a Pearson’s correlation coefficient of 0.6392 was established between changes in mRECIST and tumor volume
Brahim W. et al., 2019 [30]	After supervised delineation of thoracic cavity the tumor was automatically extracted through a statistical texture analysis approach	CT	10	To propose a diagnostic aid system capable of segmenting and measuring the pleural thickening caused by MPM	Manual segmentation of a representative database	The algorithm obtained an average Jaccard index of 0.72
Gudmundsson E. et al., 2018 [31]	Two convolutional neural networks (CNN) were trained to segmentate pleural thickenings of left and right hemithorax	CT	130	To automatically segmentate MPM on CT scans using CNNs	Manual segmentation of 8 different observers	Median DSC ranged from 0.662 to 0.800 over the two test sets
Gudmundsson E. et al., 2020 [32]	Two CNNs were trained for segmentation of tumor implementing layers pretrained on ImageNet	CT	203	To automatically segmentate MPM on CT scans using CNNs also in more complex scenarios as of pleural effusion	Manual segmentation on 2 different test sets	Median DSC of 0.69 on the tumor and a fusion test set
Armato III S. G. et al., 2015 [33]	Computation of CT-based tumor volume, after manual segmentation by a radiologist, as a number of pixels	CT	28	This study evaluated the validity of image-based tumor volume against the physical volume of the tumor bulk		A correlation coefficient r-squared value of 0.66 was found
Gill R. R. et al., 2016 [34]	Semiautomated segmentation using HU thresholding with manual editing (exclusion of pleura effusion and chest wall musculature) by 2 thoracic radiologists	CT	129	To assess feasibility and logistics of setting up a quantitative imaging study for clinical staging of MPM	AJCC pathological staging was assessed by the two radiologists on preoperative CT scans	A good overall correlation between computed tumor volume was found (Spearman Corr. = 0.822); tumor volume correlated with pathological T stage (results are reported in a separate manuscript)
Gill R. R. et al., 2018 [35]	Semiautomated segmentation with HU thresholding and manual correction to exclude pleural fluid and normal tissue; a software integrated measurement caliper was used to measure maximal fissural thickness	CT	472	To improve prognostic classification of MPM exploring alternative staging models based on quantitative parameters such as volume assessed from CT scans (VolCT) and maximal fissural thickness (Fmax)	AJCC pathological staging information were obtained from the electronic medical record for each patient	A quantitative model with both VolCT and Fmax was found to be a better prognostic classifier compared to cTNM (c-index = 0.638, p = 0.001)
Burt B.M. et al., 2020 [36]	The 3D volume feature of the software was implemented to render the thoracic cage and, after manual removing of undesired objects, to calculate TCV (thoracic cage volume)	CT	170	To determine the incidence and preoperative predictors of diffuse chest wall invasion		In univariable analysis decreased TCV demonstrated the strongest association with diffuse chest wall invasion (p = 0.009)
Brahim W. et al., 2017 [37]	A texture analysis method based on statistical approach was implemented to segmentate MPM	CT	10	To present a texture-based segmentation method of the MPM from thoracic CT scans	Tumoral regions were manually contoured and used as ground truth	The average Jaccard index was 0.73