Investigators of the Vall d'Hebron Institute of Oncology's (VHIO) Radiomics Group, led by Raquel Perez-Lopez, have developed SALSA (System for Automatic Liver Tumour Segmentation and Detection), a fully automated deep learning-driven tool for the precise and completely automated detection and monitoring of liver tumours (hepatocellular carcinoma). Results of this work have been published in Cell Reports Medicine.

Medical images, including computed tomography (CT) scans, provide professionals in oncology with essential information for cancer detection, treatment planning, and response evaluation.

"However, the precise delineation of tumours for volume analysis—tumor contouring—poses practical challenges and often causes a bottleneck in research projects and clinical applications involving volumetric disease assessment. This task is not only time-consuming but also prone to variability between and among different observers," said Perez-Lopez, corresponding author of the paper.

Primary liver cancers such as hepatocellular carcinoma (HCC) are often diagnosed in advanced stages, with limited treatment options available and an unfavourable prognosis. Furthermore, the liver is a common organ for metastases originating from other primary cancers, significantly impacting treatment outcomes.

Aimed at addressing these clinical challenges, VHIO investigators have developed SALSA, an AI-aided tool that works directly on medical images, automatically detecting and delineating liver tumours, both primary and metastatic.

"For the development and training of SALSA, we used the existing nnU-Net segmentation method and introduced data obtained from 1598 computed tomography scans of 4908 primary or metastatic liver tumours," explained Maria Balaguer-Montero, a Ph.D. Student of VHIO's Radiomics Group and first author of the study.

Showing superior accuracy for cancer detection and precise quantification of tumour burden, surpassing state-of-the-art models and the radiologists' inter-agreement, SALSA could potentially enhance cancer detection, treatment planning, and response evaluation.

"SALSA demonstrated high accuracy tumour detection with precision of over 99% at the patient level, and a lesion-by-lesion detection precision of almost 82% in the external validation cohort," added Balaguer-Montero.

"This novel deep learning-driven tool has shown precise and automated identification and delineation of liver cancer on CT images, facilitating a more precise quantification of tumour burden—a crucial factor in cancer prognosis and treatment—with no prior manual prompt requirements.

"Our validation across several tests and external cohorts highlights SALSA's effectiveness and reliability, matching and often surpassing the accuracy of expert radiologists," observed Perez-Lopez.

Imaging biomarker discovery for treatment response assessment

A key objective of VHIO's Radiomics Groups is to develop medical imaging tools based on imaging biomarkers, including tumour volume, density, and texture. Emerging technologies with artificial intelligence and machine learning techniques are rapidly advancing the field of personalised oncology by providing a more precise evaluation of treatment response in individual patients.

"Current clinical response criteria are limited. Tumour diameter is measured in 2D without considering volume and, in the case of metastases, only two tumours per organ and a maximum of five lesions per patient are evaluated," said Pérez-López.

"SALSA could be useful for the management of patients with liver cancer, enabling the measurement of parameters such as total tumour volume, density, or texture to assess response to therapy and support therapeutic decision-making to enhance patient outcomes," she concluded.