The Clinical Problem Accurately identifying and measuring all liver tumors on CT is critical for cancer staging, surgical planning, and treatment response assessment, but doing it manually is time-consuming, subjective, and varies between radiologists - making fully automated solutions transformative.
The Solution SALSA (System for Automatic Liver tumor Segmentation and Detection) is a fully automated deep learning tool developed at Vall d'Hebron Institute of Oncology (VHIO) in Barcelona. Unlike tools requiring user prompts or region-of-interest drawing, SALSA requires no human input - it processes a CT scan and outputs all liver tumor detections and delineations automatically.
The Scale SALSA was developed on 1,598 contrast-enhanced CT scans containing 4,908 liver tumors from 1,306 patients - a dataset encompassing both primary liver cancers (HCC, cholangiocarcinoma) and liver metastases from colorectal, lung, and neuroendocrine primary tumors.
Performance Benchmark SALSA achieves patient-wise tumor detection precision of 99.65% and DSC of 0.760 for tumor delineation in external validation cohorts, outperforming both the top-performing LiTS challenge model and the inter-reader agreement among expert radiologists.
Architecture Selection Three state-of-the-art neural network architectures were compared, including transformer-based models. The 3D U-Net cascade model from nnU-Net achieved the top performance and was selected as the SALSA tool - a cascade design that first segments the entire liver region then refines tumor segmentation within it.
Heterogeneous Training Data Training data included CT scans with varying acquisition protocols, scanner hardware, slice thicknesses, and image quality - intentionally mirroring the diversity of real-world clinical imaging including common artifacts like metal and motion artifacts. This heterogeneity is crucial for building a tool that generalizes beyond controlled research settings.
Four External Validation Datasets SALSA was validated against four independent external datasets: LiTS (Liver Tumor Segmentation Challenge), MSD-Hepatic Vessels (Medical Segmentation Decathlon), TCIA-CRLM (colorectal liver metastases), and TCIA-HCC-TACE-Seg (primary HCC) - covering the full spectrum of liver cancer types encountered clinically.
Radiologist Comparison A subgroup of 25 patients from the test cohort was evaluated by three blinded expert radiologists who independently delineated all tumors. Both intra-radiologist variability (same radiologist, repeated segmentation) and inter-radiologist variability were quantified, then compared to SALSA's performance.
Patient-Level Detection At the patient level - determining whether any liver tumors are present - SALSA achieved 99.65% precision and 94.17% recall in the external validation cohort. This near-perfect patient-level detection means SALSA would correctly identify as cancer-positive 94 out of 100 patients with liver tumors.
Tumor-Level Performance At the individual lesion level (detecting each tumor separately), SALSA achieved 81.72% precision and 57.92% recall in external validation. The lower lesion-level recall reflects the challenge of detecting very small lesions under 1 cm - which are also classified as non-measurable by RECIST guidelines.
Segmentation Quality Delineation quality, measured by Dice similarity coefficient (DSC), was 0.760 at the tumor level in external validation - outperforming both state-of-the-art models and the inter-radiologist agreement (DSC 0.778 and 0.736 for the two blinded radiologists respectively).
Surpassing Human Performance In the 25-patient radiologist comparison subset, SALSA achieved an F1-score of 75.89 and DSC of 0.800 - better than both blinded radiologists (F1 scores of 68.72 and 47.81) and approaching the reference expert radiologist's intra-reader performance (DSC 0.820).
Prognostic Validation SALSA's automatically quantified total liver tumor volume proved to have significant prognostic value: higher liver cancer burden was associated with poorer patient outcomes across solid tumor types (p = 0.028, hazard ratio 1.692; 95% CI 1.055-2.715).
Volume Beyond Diameter Current clinical standards (RECIST) track the diameter of up to five target lesions. Total tumor volume, automatically quantified by SALSA across all lesions, captures a more complete picture of liver disease burden than diameter-based measurements - particularly in patients with multiple lesions of varying sizes.
Implications for Response Assessment Because SALSA measures every detectable tumor rather than a selected subset, it could detect earlier and more sensitive treatment response signals - capturing overall liver disease regression even when individual target lesions change modestly.
Performance Across Tumor Types Performance was consistent across different liver tumor types (primary vs. metastatic) and CT slice thicknesses, but hypodense tumors were detected more readily than isodense lesions - which have lower contrast against the background liver parenchyma.
HCC Surveillance Enhancement For high-risk cirrhotic patients undergoing surveillance, SALSA applied to CT scans could flag the presence and location of all hepatic lesions automatically, reducing radiologist time and ensuring no lesion is overlooked in livers with complex background nodularity.
Treatment Planning Revolution Surgical planning for hepatic resection requires knowing the number, size, and location of all tumors plus the expected future liver remnant. SALSA's automated delineation of every tumor provides the comprehensive input data that surgical planning algorithms need, without the hours of manual segmentation currently required.
Clinical Trial Efficiency In clinical trials evaluating new HCC and liver metastasis therapies, accurate response measurement is central to efficacy endpoints. SALSA's volumetric assessment of total liver disease burden could serve as a more sensitive endpoint than standard RECIST, detecting earlier and more comprehensive treatment effects.
Scalable Population Screening As multi-detector CT becomes increasingly available globally, automated liver tumor detection by SALSA could be deployed as a background analysis layer on all abdominal CTs - flagging unsuspected liver lesions in patients scanned for other indications and enabling incidental HCC detection at earlier, more curable stages.
Small Lesion Detection Gap SALSA's lesion-level recall of 57.92% in external validation reflects reduced sensitivity for tumors under 1 cm - exactly the lesions that are most important to detect early for curative intervention. Improving small lesion detection is the highest priority for future development.
Non-Contrast CT Challenge SALSA was developed on contrast-enhanced CT, which highlights lesion vascularity. Performance on non-contrast CT - commonly used in surveillance settings - has not been validated and likely requires dedicated retraining.
Primary Tumor Type Breakdown While SALSA was trained on a diverse dataset, primary HCC constituted only 19 patients in the development cohort. Dedicated validation in larger HCC-specific datasets (separate from liver metastases) is needed to confirm performance for the primary liver cancer use case.
Regulatory Path SALSA is currently available for research use at radiomics.vhio.net but requires prospective multi-center trials and regulatory clearance as a software as a medical device (SaMD) before clinical deployment - a path that requires substantial additional validation data and safety evidence.