A CT-based deep learning-driven tool for automatic liver tumor detection and delineation in patients with cancer

Abstract

Liver tumors, whether primary or metastatic, significantly impact the outcomes of patients with cancer. Accurate identification and quantification are crucial for effective patient management, including precise diagnosis, prognosis, and therapy evaluation. We present SALSA (system for automatic liver tumor segmentation and detection), a fully automated tool for liver tumor detection and delineation. Developed on 1,598 computed tomography (CT) scans and 4,908 liver tumors, SALSA demonstrates superior accuracy in tumor identification and volume quantification, outperforming state-of-the-art models and inter-reader agreement among expert radiologists. SALSA achieves a patient-wise detection precision of 99.65%, and 81.72% at lesion level, in the external validation cohorts. Additionally, it exhibits good overlap, achieving a dice similarity coefficient (DSC) of 0.760, outperforming both state-of-the-art and the inter-radiologist assessment. SALSA's automatic quantification of tumor volume proves to have prognostic value across various solid tumors (p = 0.028). SALSA's robust capabilities position it as a potential medical device for automatic cancer detection, staging, and response evaluation.

Keywords: deep learning; delineation; detection; imaging; liver tumors.

Graphical abstract

Figure 1

Overview of the study population and design (A) Distribution of the data included in the study. It details the number of CT scans and liver tumors involved in the development and test cohorts as well as the collected external validation cohorts. (B) Overview of the study workflow and methodological framework, including the tested architectures and evaluation metrics for per-patient and tumor-wise assessments. It features a schematic representation of two side studies: (1) benchmarking against professional radiologists through intra- and inter-reader variability studies and (2) expert preference analysis comparing manual (ground truth) segmentations with those generated by SALSA. Additionally, it explores the use of automated liver tumor quantification as a prognostic biomarker in patients with cancer.

Figure 2

SALSA’s performance evaluation (A and B) Delineation performance evaluated by the intersection of ground truth masks with those automatically generated by SALSA, calculated using the dice similarity coefficient (DSC). Results from the three tested architectures in the test set (orange) and external validation set (blue) are shown for both patient-wise (A) and tumor-wise (B) levels. (C) Evaluation of various metrics for detection (precision, recall, and F1-score) and delineation (DSC and Jaccard Index [JI]) performance by SALSA. (D and E) Analysis of the impact of tumor density (D) and volume (E) on delineation performance, highlighting SALSA’s reduced efficacy in hyperdense and small liver tumors. Significance was calculated using independent two-sample t tests with Bonferroni correction and is represented as: ∗p < 0.05; ∗∗p ≤ 0.01; ∗∗∗p ≤ 0.001; ∗∗p ≤ 0.0001; NS, not significant. Data are represented as median and interquartile range (IQR) in (A), (B), and (E). See also Tables S3, S4, and S6.

Figure 3

Visual comparison of the automatically delineated contours with radiologist-generated ground truth Representative cases of liver tumors delineated by SALSA (red lines) compared to the ground truth (blue lines) segmented masks. Yellow dashed boxes in (A) and (C) indicate the regions magnified in (B) and (D) for improved visualization. The delineations are displayed as colored masks, highlighting areas of agreement and discrepancies between the assessments.

Figure 4

Benchmarking SALSA against state-of-the-art models and inter-radiologist metrics (A) SALSA benchmarks against the top-performing LiTS model for both detecting and delineating liver tumors. Metrics include detection (precision, recall, and F1-score) and delineation (dice similarity coefficient [DSC] and Jaccard Index [JI]) across the test set using patient-wise and tumor-wise approaches. (B) Comparison of tumor delineation overlaps. Overlap with ground truth delineations is compared for segmentations by the same radiologist on two occasions (Rad 1), two independent radiologists (Rad 2 and Rad 3), and the SALSA models (nnU-Net, Vanilla U-Net, and TransUNet), providing a benchmark of model performance against human experts. (C) Radiologist preferences for manual versus automated liver tumor delineation. Three expert radiologists (Rad 1, Rad 2, and Rad 3) assessed their preferences between manual segmentations (performed by an expert radiologist) and automated segmentations (SALSA), with the option to express no preference. Data are represented as median and interquartile range (IQR) in (A) and (B). See also Tables S5 and S9.

Figure 5

Total tumor volume quantification as a prognostic biomarker Kaplan-Meier curves and log rank test results for overall survival are shown for 141 patients in the test set (A) and 197 patients from the TCIA-CRLM dataset in the external validation cohort (B). Patients were grouped by thresholding total liver tumor volume, using both ground truth (left) and SALSA-generated volumes (right), at the median value for each cohort, demonstrating SALSA’s potential for biomarker research.