PREOPERATIVE DETECTION OF HEPATOCELLULAR CARCINOMA’S MICROVASCULAR INVASION ON CT-SCAN BY ARTIFICIAL-INTELLIGENCE AND RADIOMICS ANALYSES

Introduction. Microvascular invasion (MVI) is the main risk factor for overall mortality and recurrence after surgery for hepatocellular carcinoma (HCC). Its diagnosis can be made only postoperatively on the histological specimen. The aim of this preliminary study was to train machine-learning models to predict MVI on preoperative CT scan (fig.1).
Methods. Clinical data and 3-phases CT scans were retrospectively collected among 4 Italian centres. DICOM files were manually segmented to detect the liver and the tumor(s). An already available segmentation algorithm (Nnunet) was retrained to obtain automatic detection focused on HCC. An implementation was added to automatically extract radiomics features from the tumoral, peritumoral (among 5mm from the tumor margin) and healthy liver areas in each phase. Performance comparison between manual and algorithm segmentations was measured by intersection over union (Jaccard Index). Data obtained were explored and principal component analysis (PCA) was performed to reduce the dimensions of the dataset, keeping only the PC’s explaining 95% of the variability. After normalization, data were divided between training (70%) and test (30%) sets. Random-Forest (RF), fully connected MLP Artificial neural network (neuralnet) and extreme gradient boosting (XGB) models were fitted to predict MVI. Hyperparameters tuning was made per each model to reduce the out-of-bag error. Prediction accuracy was estimated in the test set and employed as the study end-point.
Results. Between 2008 and 2022, 218 consecutive preoperative CT scans of patients affected by HCC and submitted to surgery were collected with the relative clinical data. At the histological specimen, 72 (33.02%) patients had MVI. The Jaccard index between manual and algorithm segmentations was overall 90%. First and second order radiomics features were extracted, obtaining 672 variables per patient. After data exploration, PCA selected 58 dimensions explaining >95% of the variance. After standardization and normalization, RF, neuralnet and XGB were fitted to predict the presence of MVI. Tuning parameters were: 1) RF: n.tree=500, mtry=30; 2) Neuralnet: 2 hidden layer with 40 and 20 neurons, learning rate= 0.001, threshold for termination= 1%, activation function= sigmoid; 3) XGB: nrounds = 100, max_depth = 3, eta = 0.3. The models were then fitted in the testset to estimate prediction accuracy by confusion-matrix. RF was the best performer (Acc=98.4%, 95%CI: 0.91-0.99, Sens: 95.2%, Spec: 100%, PPV: 100% and NPV: 97.7%, fig.2).
Conclusion. Our model allowed an impressive prediction accuracy of the presence of MVI at the time of HCC diagnosis, never reached until now. This could lead to change the treatment allocation, the surgical extension and the follow-up strategy for those patients. The algorithm will be freely distributed online for medical purpose.