EUS-based intratumoral and peritumoral ML radiomics for distinguishing insulinomas from non-functional PNETs.

Front Endocrinol (Lausanne) 2024 AI 6 Explanations View Original
Original Paper (PDF)

Unable to display PDF. Download it here or view on PMC.

Plain-English Explanations
Pages 1-2
Distinguishing Two Types of Pancreatic Neuroendocrine Tumors

Pancreatic neuroendocrine tumors (PNETs) are rare tumors that arise from hormone-producing cells in the pancreas, accounting for 1-3% of all pancreatic tumors. They are broadly divided into functional PNETs (F-PNETs), which secrete hormones causing symptoms, and non-functional PNETs (NF-PNETs), which typically remain silent until they reach an advanced stage.

Insulinomas are the most common type of F-PNET. They cause recurrent hypoglycemia (dangerously low blood sugar) because they continuously secrete insulin. Most insulinomas are benign, but their symptoms are often mistaken for neurological disorders, leading to lengthy misdiagnosis. NF-PNETs, by contrast, are typically larger at diagnosis, more aggressive, and carry significantly worse prognoses with higher rates of lymph node invasion and liver metastasis.

Distinguishing insulinomas from NF-PNETs before surgery is critical because their management differs substantially - most guidelines recommend immediate surgery for insulinomas, while small NF-PNETs under 2 cm may be safely observed. Current imaging tools like CT have a 31% miss rate for insulinomas. Endoscopic ultrasonography (EUS) detects all lesions but relies mainly on visible anatomical features, limiting its specificity without radiomics enhancement.

TL;DR: Insulinomas and NF-PNETs require different management strategies, but distinguishing them preoperatively is challenging because they look similar on standard imaging.
Pages 3-5
EUS Radiomics Extracted from Both Inside and Around the Tumor

A total of 106 patients - 61 with insulinomas and 45 with NF-PNETs - were enrolled and split 7:3 into training (n=74) and test (n=32) cohorts. All patients had confirmed diagnoses by surgical pathology or EUS-guided fine-needle aspiration. EUS images were acquired with standardized protocols by a specialist with over 12,000 EUS procedures.

Two EUS specialists independently traced the intratumoral region of interest (ROI) using ITK-SNAP software. The peritumoral ROI was created by expanding the intratumoral boundary outward by 3 mm, capturing the tissue immediately surrounding the tumor. This peritumoral zone reflects the tumor microenvironment - the biological context in which the tumor grows - which may carry independent diagnostic information beyond the tumor itself.

A total of 107 radiomics features were extracted from both intratumoral and peritumoral regions, including 18 first-order statistical features, 14 shape features, and multiple texture features from GLCM, GLRLM, GLSZM, and NGTDM matrices. After Mann-Whitney U filtering, Spearman correlation pruning, and LASSO regression, 4 intratumoral, 6 peritumoral, and 5 combined features with nonzero coefficients were retained.

TL;DR: 107 quantitative features were extracted from the tumor core and the 3 mm surrounding tissue ring in 106 PNET patients' EUS images, then filtered down to the most discriminative features.
Pages 6-8
LightGBM Wins - And Peritumoral Data Adds Significant Value

Six machine learning algorithms were compared on the intratumoral features: Logistic Regression, Random Forest, XGBoost, LightGBM, Extra Trees, and Multilayer Perceptron. Random Forest, Extra Trees, and XGBoost showed signs of overfitting. The LightGBM model demonstrated the best balance between training and test performance, achieving an AUC of 0.879 in training and 0.750 in the test cohort as the intratumoral radiomics model.

A peritumoral model was then built using the same LightGBM algorithm. Strikingly, the peritumoral model's AUC in the test cohort (0.750) matched the intratumoral model's performance exactly (p=1.000 on the DeLong test), demonstrating that the tissue surrounding the tumor contains as much diagnostic information as the tumor itself. This supports the view that the tumor microenvironment is diagnostically informative for PNET typing.

The combined radiomics model integrating both intratumoral and peritumoral features achieved the highest performance: AUC of 0.876 in training and 0.835 in the test cohort (95% CI: 0.698-0.973). The DeLong test confirmed the combined model significantly outperformed the intratumoral model alone in the test set (p=0.045), validating the benefit of including peritumoral information.

TL;DR: LightGBM outperformed other algorithms, and the combined intratumoral plus peritumoral model achieved an AUC of 0.835 - significantly better than either region alone.
Page 10
A Nomogram Combining Radiomics and Tumor Diameter

The only clinical feature significantly different between insulinomas and NF-PNETs was tumor diameter - insulinomas were substantially smaller (mean 13.7 mm) compared to NF-PNETs (mean 33.9 mm in training, 28.7 mm in test). All other macroscopic EUS features (shape, margin, echogenicity, calcification, location) showed no significant differences, confirming that diameter is the only standard clinical discriminator.

A nomogram was constructed combining the combined radiomics signature and tumor diameter using logistic regression. This visual tool assigns points to each factor and sums them to produce a final probability score for insulinoma vs. NF-PNET. The nomogram achieved an AUC of 0.929 (95% CI: 0.846-0.984) in the training cohort and 0.913 (95% CI: 0.794-0.992) in the test cohort - substantially outperforming either the radiomics signature or diameter alone.

Calibration analysis showed that the nomogram's predicted probabilities closely matched actual observed outcomes, with a mean absolute error of only 0.024. Decision curve analysis confirmed that the nomogram provided clinical net benefit across the entire high-risk probability range (0 to 1.0), outperforming both the treat-all and treat-none approaches in terms of clinical utility.

TL;DR: Adding tumor diameter to the radiomics signature in a nomogram boosted AUC to 0.929 in training and 0.913 in testing, providing a practical and highly accurate clinical decision tool.
Pages 11-12
Why EUS Radiomics Succeeds Where Visual Inspection Fails

Standard EUS relies on macroscopic features like echogenicity, margin clarity, and shape to characterize pancreatic lesions. This study confirmed that none of these visual features differed significantly between insulinomas and NF-PNETs - both tumor types appear nearly identical to the naked eye. Radiomics features extract quantitative textural patterns at the pixel level that human observers cannot perceive, capturing deeper biological heterogeneity.

The clinical importance of the peritumoral ROI is a key finding. Prior research in other cancers has shown that the peritumoral region reflects biological interactions between the tumor and its local environment - including immune cell infiltration, stromal remodeling, and vascular changes. For PNETs specifically, the authors propose that peritumoral radiomics features may encode the degree of local invasion and hormone-secretory activity that influences the surrounding tissue differently in insulinomas vs. NF-PNETs.

The study is limited by its retrospective single-center design and relatively small sample size (106 patients for a rare tumor type). Insulinoma diagnosis in some patients relied on clinical and biochemical criteria rather than pathology alone. Future studies should enroll larger multi-center cohorts to validate the nomogram's diagnostic performance and test its generalizability across different clinical settings and EUS equipment.

TL;DR: Radiomics extracts diagnostic information invisible to human observers, and peritumoral tissue surrounding the tumor provides independent biological signals that improve PNET classification.
Pages 1, 12
First EUS Radiomics Model for Distinguishing PNET Subtypes

This study is the first to apply EUS-based intratumoral and peritumoral radiomics with machine learning to distinguish insulinomas from NF-PNETs. The approach addresses a genuine clinical gap: these two tumor types look similar on conventional imaging, yet require completely different management strategies with meaningfully different prognoses.

The combination of the LightGBM-based radiomics signature and tumor diameter in a nomogram provides a clinically practical tool that achieved an AUC of 0.913 on independent test data. This level of performance, using only EUS images and a simple measurement, makes the tool suitable for implementation in centers that perform routine EUS for pancreatic evaluation.

Beyond diagnosis, the framework established here - extracting and analyzing both intratumoral and peritumoral radiomics features - represents a broadly applicable methodology that can potentially be adapted to other challenging differential diagnosis problems in pancreatic oncology, such as distinguishing pancreatic ductal adenocarcinoma from autoimmune pancreatitis or different grades of PNETs.

TL;DR: The first EUS radiomics model for PNET subtype differentiation achieved an AUC of 0.913 using a nomogram combining radiomics signature and tumor diameter, filling a significant clinical diagnostic gap.
Citation: Open Access, 2024. Available at: PMC11215175.