Tumor Voxel Dose-Response Matrix Prediction Using Deep Learning

PURPOSE/OBJECTIVE(S)

Tumor voxel dose-response matrix (DRM) can be assessed using a series of FDG-PET/CT feedback images acquired during radiotherapy. Predicting the tumor voxel DRM earlier is crucial for effectively implementing adaptive treatment management. However, it is also challenging due to FDG uptake dynamic fluctuation in tumor cells. This study investigated the feasibility of predicting tumor voxel DRM during the early treatment weeks using the advanced deep learning (DL) technique.

MATERIALS/METHODS

Serial FDG-PET/CT images were acquired at the pretreatment (pre-Tx), the 2nd and 4th treatment weeks during standard chemo-radiotherapy (35 × 2 Gy) from each of the 50 patients with head and neck squamous cell carcinomas (HNSCC). The reference value of tumor voxel DRM (DRMref), representing the average metabolic change ratio during the treatment, was determined using a linear regression performed on the standard uptake values (SUV)s obtained at the pre-Tx (SUV0), the 2nd (SUV2) and the 4th (SUV4) treatment weeks following deformable PET/CT image registration. A DL model, 3D residual-Unet with a total of 3.4 million parameters, was trained to predict the tumor voxel DRMref with using the SUV0 and SUV2 matrices as inputs. The performance of the DL model was evaluated using 10-fold cross-validation and was compared to that of a linear regression (LR) model determined on the SUV0 and SUV2 matrices.

RESULTS

The mean (SD) of the tumor voxel DRMref was 0.46 (0.2) over all 34612 tumor voxels. The predicted tumor voxel DRM was 0.5 (0.38) and 0.46 (0.15) for the LR model and the DL model, respectively. For those resistant voxels (23.7% of all tumor voxels) with a DRMref > 0.6, the DRM deviation was 0.13 (0.4) and -0.11 (0.13) for the LR model and the DL model, respectively. For those sensitive voxels (76.3%) with a DRMref ≤ 0.6, the DRM deviation was 0.01 (0.23) and 0.03 (0.08) for the LR model and the DL model, respectively.

CONCLUSION

The proposed DL model can predict the tumor voxel DRM with a single FDG-PET feedback image acquired during the 2nd treatment week of radiotherapy for HNSCC patients. The prediction accuracy was improved compared to that of the LR model with a substantial reduction in the variances of the prediction errors. This work demonstrates the great potential of utilizing DL techniques to improve the efficiency of tumor response assessment and adaptive treatment management.

Challenging the Role of Subtotal Resection Following Proton Radiotherapy for Adenoid Cystic Carcinoma of the Head and Neck

PURPOSE/OBJECTIVE(S)

To report long-term outcomes of patients with adenoid cystic carcinoma (ACC) of the head and neck treated with proton radiotherapy.

MATERIALS/METHODS

On this IRB-approved, single institutional prospective outcomes registry, 56 patients were included with de novo, nonmetastatic adenoid carcinoma of the head and neck treated with primary (n = 9) or adjuvant proton therapy from June 2007 to December 2021. The cohort had 30 women and 26 men with a median age of 57 years (range, 10-81 years). Twenty-nine percent (n = 16) had intracranial extension, 23% (n = 13) had orbital extension, and 55% (n = 31) had clinical cranial nerve involvement at the time of radiotherapy. Thirty patients underwent gross total resection (GTR), 26 had gross disease at the time of treatment undergoing a subtotal resection (STR) (n = 17) or biopsy alone (n = 9). The median dose to the primary site was 72.6 GyRBE (range, 64-74.4 GyRBE) delivered in either once (n = 19) or twice (n = 37) daily treatments. Thirty patients received either elective nodal irradiation (ENI) in a node negative neck or concurrent chemotherapy.

RESULTS

With a median follow-up of 6.2 years (range, 0.9 - 14.7 years), the 5-year local-regional control (LRC), disease free survival (DFS), cause-specific survival (CCS) and overall survival (OS) were 88%, 85%, 89%, and 89%, respectively. Cranial extension (p = 0.003) and gross residual tumor (p = 0.0388) were factors associated with decreased LRC. While LRC for those with a GTR was 96%, those with STR or biopsy alone were 81% and 76%, respectively. T-stage (p = 0.0154), cranial extension (p = 0.0056), extent of resection (p = 0.0355), and gross residual tumor (p = 0.0094) were associated with decreased DFS. T-stage (p = 0.0099), extent of surgery (p = 0.029) and gross residual tumor (p.0071) were associated with decreased CCS. The 5-year cumulative incidence of clinically significant late grade ³3 toxicity was 15% and the crude incidence at most recent follow-up was 23% (n = 13). There was no LRC benefit with ENI (p = 0.94).

CONCLUSION

Proton therapy provides excellent disease control for head and neck ACC with acceptable toxicity. Gross residual disease at the time of treatment and intracranial involvement were significant prognostic features for worse outcomes. STR did not confer benefit over biopsy only at 5-years and may question the role of extensive and morbid operations if GTR is not technically feasible.

Deep Learning-Based Dose Prediction Model for Automated Spot-Scanning Proton Arc Planning

PURPOSE/OBJECTIVE(S)

Spot-scanning proton arc (SPArc) is a novel technique that employs a planning optimization algorithm to select the energies and positions of spots along a dynamic rotational arc trajectory. The SPArc technique has the potential to achieve superior dose conformality and treatment delivery efficiency over intensity-modulated proton therapy. However, creating such a SPArc plan using existing approaches is time-consuming and computationally extensively. This study investigated the feasibility of using the deep learning (DL) technique to predict the 3D dose distribution of the SPArc treatment plan, leveraging the prior knowledge acquired from conventional intensity-modulated radiation therapy (IMRT) plans.

MATERIALS/METHODS

A DL model, 3D-Unet with residual connections and attention gates, was trained using an open-source database of CT images, critical structures, and IMRT plans from 340 head and neck cancer patients (HNC) as the base model. Transfer learning technique was applied to fine-tune the model parameters using the SPArc treatment plans created on the HNC patients from an in-house dataset, where the SPArc treatment plans (including control point sampling, energy layer distribution, arc trajectory, etc.,) were optimized using a previously developed iterative approach. The performance of the DL model was evaluated by comparing predicted and planned doses over 17 SPArc treatment plans by using 4-fold cross-validation.

RESULTS

The SPArc planning time per patient was 8∼12 hours, while the dose prediction time was reduced to 2∼3 minutes using the proposed DL model. The deviation of D95 in the target was (-1.8±1.6) %. The deviation of the mean dose in the parotids, cord, mandible, and brainstem were (2.5±6.5) %, (-0.5±4.3) %, (1.4±3.9) %, and (3.4±8) % of the prescription, respectively. The dice similarity coefficients of the 80%, 70%, and 60% isodose lines were (0.9±0.09), (0.93±0.01), and (0.94±0.01), respectively.

CONCLUSION

Our results demonstrate that a DL-based dose prediction model can be created with a limited number of SPArc treatment plans through transfer learning. The DL model can directly predict the 3D dose distribution in minutes for automated planning. This study paves the roadmap to develop a quick clinical decision platform for the optimal selection among the multi-treatment modalities.

Patterns of Recurrence for Lymph Node Positive Cutaneous Melanoma in the Era of Immunotherapy

PURPOSE/OBJECTIVE(S)

The management of patients with lymph node positive, nonmetastatic cutaneous melanoma has changed significantly with the adoption of anti-PD-1 immunotherapy. We compared clinical outcomes and patterns of recurrence of patients with lymph node positive, nonmetastatic cutaneous melanoma who received adjuvant immunotherapy with those who did not receive immunotherapy.

MATERIALS/METHODS

Patients with lymph node or in-transit metastasis positive, nonmetastatic cutaneous melanoma diagnosed from 2013-2020 were identified from a prospectively-maintained, single-institution database. Baseline patient demographics and treatment details were recorded. Patients were stratified by receipt of adjuvant immunotherapy. Local and regional recurrence, distant metastasis, first site of recurrence, and overall survival (OS) were recorded, and rates of OS, progression-free survival (PFS), freedom from distant metastasis (ffDM), local recurrence (ffLR), and regional lymph node recurrence (ffRR) were estimated using the Kaplan-Meier method. Wilcoxon rank-sum test, Fisher's exact test, and chi-square test of independence were used to analyze variables between groups.

RESULTS

Ninety-two patients were included with a median age of 67 years and a median follow-up of 2.9 years. Fifty-six (61%) patients received adjuvant immunotherapy for a median duration of 8.5 months. Fifty-three (58%) patients had a sentinel lymph node dissection alone while thirty-two (35%) had a complete regional lymph node dissection. A median of 3 lymph nodes were removed (range 0-83) with a median of 1 positive lymph node (range 1-15). Twenty-four (26%) patients had extranodal extension, thirty-nine (42%) had ulceration of the primary site, and twenty-eight (30%) had LVSI present. Median age of patients receiving immunotherapy versus those who did not was 63 years versus 72 years (p = 0.02). There were no differences in primary site, disease thickness, Clark stage, primary site ulceration, LVSI, extranodal extension, satellite metastasis, AJCC T- and N-stage, and completion lymph node dissection rates between the groups. Patients receiving immunotherapy had significantly improved OS (HR = 0.3, p = 0.001) with no difference in PFS (HR = 0.6, p = 0.08), ffDM (HR = 0.66, p = 0.28), ffLR (HR = 0.18, p = 0.1), and ffRR (HR = 2.5, p = 0.08). The initial site of recurrence was regional lymph nodes in thirteen (23%) patients receiving immunotherapy versus two (6%) patients who did not receive immunotherapy (p = 0.01). On UVA, male gender, number of positive lymph nodes, clinically positive lymph nodes, extranodal extension, in-transit metastasis, LVSI, and tumor thickness were significant predictors of regional lymph node recurrence.

CONCLUSION

Patients with lymph node positive cutaneous melanoma receiving immunotherapy had a higher rate of initial disease progression in the regional lymph nodes compared with patients not receiving immunotherapy.

Mitigation of Dosimetric Uncertainty in MRI-Based Proton Planning Using Spot-Scanning Proton Arc (SPArc) Technique

PURPOSE/OBJECTIVE(S)

MRI-based synthetic CT (SCT) images created using generative adversarial network (GAN) have been demonstrated to be feasible for intensity-modulated proton therapy (IMPT) planning. However, dose calculation accuracy can be uncertain in some regions within/near the target of head and neck patients due to the local CT number estimation error or sharp dose fall-off. This study investigated the feasibility of using the SPArc technique to mitigate such dosimetric uncertainty.

MATERIALS/METHODS

A GAN using a 3D U-net as the generator and a 6-layer 3D convolutional neural network as the discriminator was trained with T1-weighted MR-CT image pairs from 162 nasopharyngeal carcinoma patients (14 for validation). The generator was used to generate SCT images from MR images for 7 test patients. For each test patient, the CT image was used to create a SPArc plan and an IMPT plan with the same clinical objectives. The SPArc plans (control point frequency sampling, arc trajectory, etc.) were optimized using a previously developed iterative approach. The dose distributions of both SPArc plans and IMPT plans were re-calculated on the SCT images and compared to the one calculated on the CT images. The dosimetric uncertainty was quantified using the gamma index.

RESULTS

The 2%/2mm and 3%/3mm passing rates for SPArc plans were (96.9¡À2.7) % and (98.6¡À1.5) %, while the passing rates for IMPT plans were (94.0¡À3.9) % and (96.4+2.9) %. A significant reduction in dosimetric uncertainty was identified for SPArc plans (p ¡Ü0.021). Table 1 shows the passing rates for the 7 test individuals.

CONCLUSION

SPArc can mitigate the uncertainty of dose calculation in MRI-based proton planning. Further research needs to validate these findings on a larger patient cohort. The study paves the road map for using MRI for SPArc planning.