Radiomics is a science which combines radiology, mathematics and artificial intelligence techniques, using data-characterisation algorithms and mathematical analysis to extract many aspects and traits from radiological imaging; it performs a quantitative approach to diagnostic imaging acquisitions, as opposed to the classical qualitative approach made by physicians [4]. Radiomics evaluates the so-called image biomarkers as digital image texture, consisting of the single pixels and their relationship to the other pixels, as well as intensity and tissue density spatial distribution; these traits, known as radiomic features, may reveal tumoral patterns and characteristics that are not visible to the human eye. The various phases to obtain radiomics features consist of data collection, target lesion segmentation, image biomarkers detection and extraction from image texture, modeling, processing, and validation [5]. The unique imaging characteristics of different disease forms may be helpful in predicting the prognosis and therapeutic response of different types of cancer, thereby offering important information for individualized treatment. The most cutting-edge uses of radiomics are found in radiology and oncology, and therefore, in IO.

DL systems based on convolutional neural networks (CNNs) have shown potential to revolutionize the process of radiological diagnosis, increasing sensitivity in classifying neoplastic lesions, and giving the radiologist the ability to interpret, check and validate the results [6,7,8]. AI plays an important role in IO, helping physicians to achieve a higher accuracy in the diagnosis of lesions and, thus, to choose the best approach to treatment, personalized for every patient and every neoplastic lesion.

In their study, Hamm et al. tried to develop and validate a DL system based on CNNs, which classifies common hepatic lesions on multi-phasic magnetic resonance imaging (MRI), and compared its performances with those of diagnostic radiologists; test set performance in a single run of random unseen cases showed an average 90% sensitivity and 98% specificity, while the average sensitivity and specificity on these same cases for radiologists was 82.5% and 96.5%, respectively. Results showed a 90% sensitivity for classifying hepatocellular carcinoma (HCC) compared to 60–70% for radiologists [6]. Moreover, the authors have attempted to develop a proof-of-concept “interpretable” DL prototype that justifies the aspects of its predictions from the pre-trained hepatic lesion classifier, identifying and scoring radiological features. This method enables radiologists to interpret elements of decision-making behind classification decisions; this way, clinicians can validate these features by using feature maps or similar interpretability techniques and can check whether the system has accurately identified the lesion’s features in the correct locations [7]. Yasaka et al. in their clinical retrospective study investigated whether different types of liver masses could be differentiated at dynamic contrast agent-enhanced CT by using models based on DL with a CNN. Masses were diagnosed according to five categories (category A, classic HCC; category B, malignant liver tumors other than classic and early HCCs; category C, indeterminate masses or mass-like lesions—including early HCCs and dysplastic nodules—and rare benign liver masses other than hemangiomas and cysts; category D, hemangiomas; and category E, cysts). DL with CNN showed high diagnostic performance; median accuracy of differential diagnosis of liver masses for test data was 0.84 and median area under the receiver operating characteristic (ROC) curve for differentiating categories A-B from C-E was 0.92 [8]. The great promise of AI in interventional oncology is to bring precision medicine at its finest, to the level of the individual patient, through a better understanding and definition of the target lesion. Budai et al. retrospectively constructed a radiomics-based model to diagnose histotypes of renal cell carcinoma (clear-cell versus other histotypes) evaluating the CT scans of 209 patients with renal cell carcinoma, obtaining an accuracy of 78%, sensitivity of 80%, and specificity of 74%, respectively; these results were compared to the ones achieved by an expert radiologist (accuracy of 79%, sensitivity of 84%, and specificity of 69%) [9].

Staging and outcome prediction is mandatory to address the best treatment for every patient, and this also applies to IO in which locoregional treatments greatly vary based on tumor staging and patient’s outcome, particularly for HCC patients [10].

Most of the current staging systems (e.g., the Tumor, Node, Metastasis (TNM) classification) have limitations in terms of patient prognosis. On the other side, nomograms represent a useful tool in personalized care of oncologic patients, used by an increasing number of cancer centers to improve clinical practice. Nomograms are getting more and more useful in oncological settings as they are capable of estimating the personalized patient risks and prognosis based on disease and subjective characteristics. Even though nomograms’ diffusion dates back to before AI’s, artificial intelligence plays a great role, integrating prognostic and determinant variables, generating a personalized and individualized probability of a patient’s clinical event. Nomograms can be obtained for virtually all types of cancer and can predict various outcomes (prognosis at the time of diagnosis, post-treatment recurrence risk, procedure-specific survival outcomes).

AI-based nomograms with easy-to-use digital interfaces allow for fast and accurate data computation, obtaining clearer and easier to understand prognoses compared to other staging systems, improving the decision-making process. Gupta et al. retrospectively investigated CT images texture to predict grading and survival of 38 patients with suspected gallbladder neoplasm [11]. Multiple authors used radiomics-based models to predict lymph node metastases in various cancer typer (e.g., gastric, breast, bladder, colo-rectal) [12,13,14,15]. AI-based tools can also be used for the staging of the primitive tumor, as well as for prediction of aggressive disease progression [16,17,18,19]. These studies demonstrate how AI outperforms “classical” staging systems, and better determines tumor stage and presence of metastasis, granting a more accurate, personalized and faster treatment choice.

ML and DL models can be used in IO practice in predicting response to treatment of cancer patients undergoing locoregional (percutaneous or intra-arterial) therapies. Artificial neural networks that are multilayered or “deep” are the basis of DL. The various neural layers between input and output provide the DL with its plasticity and the ability to define novel patterns of intelligent classification, simulating the workings of the human brain. When compared to a human reader, who can only detect and use a portion of the total big information content of digital images, DL can automatically distinguish the pertinent features from data, allowing it to learn new patterns and determine more complex relationships. This is particularly true in IO where DL-models are capable of integrating multiple patient and tumoral variables, unseen at the human eye, to to guide clinical choices and to predict outcomes of locoregional treatments as transarterial chemoembolization (TACE), radioembolization (TARE), and percutaneous ablation.

Computed-aided detection (CAD) is an increasingly utilized tool to perform an adjunctive, second read of diagnostic imaging acquisitions (X-ray, CT, MRI) in order to assist the radiologist in the detection of pathologic lesions and improve their accuracy. Mostly used on chest X-rays and lung CT scans for pulmonary nodules [31,32]. Lee et al. evaluated the efficacy and clinical usefulness of lung nodules’ CAD in patients with colorectal cancer oligometastases to the lung, obtaining good sensitivity and specificity values [31]. Li et al. demonstrated how the recent technological advancements of CAD recognition algorithms increased the accuracy of lung nodule detection up to 99.56%, with a sensitivity of 99.3%, greatly reducing false negatives and missed detections [32]. Ahn et al. determined the usefulness of evaluating breast MRI with a CAD software in the prediction of invasive neoplasm in patients with ductal carcinoma “in situ”, to select patients for sentinel lymph-node biopsy [33]. Takamoto et al. validated a recently developed software of AI-assisted CT-based virtual hepatectomy in patients affected by liver cancer, with a focus on processing time, obtaining reliable and accurate volumetries with a significatively (p < 0.001) shorter processing time for AI-assisted reconstructions [34]. AI-assisted CNN-based virtual segmentation can also be useful and time-sparing for volumetries prior to transarterial radioembolization procedures, as demonstrated by Chaichana et al. [35]. The authors developed an automated CNN-based method for target lesion and organ segmentation on Single-Photon Emission Computed Tomography (SPECT) images obtained after 99mTc-labeled Macroaggregated Albumin (MAA) administration, obtaining a time-sparing (about 1 min per patient) and accurate segmentation method. CAD can be an extremely valuable tool for interventional radiologists, as it can provide a faster and more efficient way to detect small or barely visible lesions, leading not only to a correct and early diagnosis but also to a shorter time-to-treatment for patients, which can eventually lead to a better prognosis. As previously stated, CAD may also help IO practitioners to reduce the planning time, as in the case of pre-procedural hepatic volumetric assessment.

3. Digital Health

3.1. Virtual Reality and Augmented Reality

3.2. Robotics-Assisted Ablation

3.3. Virtual Multidisciplinary Tumor Board