Integrating deep-learning based segmentation and ‘fast’ voxel-level dosimetry tools to enable automated real-time patient specific dosimetry

Purpose: The proliferation of molecular radiotherapies has generated a clinical need for fast and accurate information that elucidates the effect of the therapy. Experience from external radiotherapy and Y-90 microsphere therapy suggests that dose-response relationships exist, but the difficulty of achieving fully patient-specific dosimetry estimates has limited our understanding for 177Lu-DOTATATE. In particular, manual 3D segmentation of organs at risk (OARs) on all timepoints (TPs) and patient-specific dosimetry calculations can require substantial effort. In this work, we exploit recent advances in machine learning and computational power to construct and evaluate a highly automatic process for patient-specific voxel dosimetry.

Methods: The automated dosimetry workflow was tested on multi-TP (2 to 4) SPECT/CT imaging data acquired post-177Lu-DOTATATE in 15 patients. All steps in the process were automated except tumor segmentation done by a radiologist. Convolutional neural networks (CNNs) were used for auto-segmentation of kidneys and liver on the CT of a reference SPECT/CT TP, while automated HU threshold-based methods were employed to segment the lungs and bone (considered to be representative of bone marrow for dosimetry). The sequential SPECT/CT images and volumes of interest (VOIs) for the reference TP were then fed into a dosimetry process that automated VOI propagation, dose rate calculation, and dose rate curve fitting and integration. VOI propagation from the reference SPECT/CT onto the other TPs was performed by an image alignment algorithm that spliced together multiple local intensity-based SPECT registrations. Conversion of aligned SPECT images into voxelized dose rate maps was performed with (a) the Dose Planning Method (DPM) ‘fast’ Monte Carlo (MC) code, benchmarked for molecular radiotherapy, and with (b) voxel S value (VSV) kernel convolution with CT-based density correction, where a long-range kernel was generated from DPM. Finally, the dose rate curves from each method were fitted with an automatic best-fit exponential selection and integrated to generate absorbed dose maps.

Results: For kidney and liver VOIs generated with the CNN, 25/45 (56%) needed no manual adjustment, and overall manual adjustment required on average 2.6min total. No more than 1min was spent cleaning the automated bone VOI, and only rarely more than 1min was spent on the lung VOIs. The automated dosimetry workflow took an estimated 15-20min to run from start to finish including both DPM and VSV methods. In particular, the runtime for simulating 1E8 nuclear decays with DPM Monte Carlo was ~2min for each TP. (All reported times are on a desktop computer with a 3.2GHz processor.) Mean absorbed doses using DPM and VSV were in excellent agreement for all VOIs except lungs: the average absolute difference across kidneys, non-tumoral liver, and bone was 2.1% (range: [0.0%, 4.4%]), while in tumors it was 1.6% (range: [0.0%, 7.5%]).

Conclusions: A highly practical process for patient-specific dosimetry has been constructed and evaluated. In this process, manual refinement of all automatic VOIs took ~4.5min, and this was only required on 1 SPECT/CT for each patient due to a VOI propagation technique. It was demonstrated in a proof-of-concept test that the 15-20min dosimetry runtime could be run entirely as a background process while other manual work was performed, meaning ~4.5min of manual time total would be required to achieve fully patient-specific absorbed doses for OARs. Studies to automate tumor segmentation using gradient-based and deep learning methods are ongoing. The rapid speed of DPM running single-threaded implies the attainability of high-precision MC dosimetry. Finally, it is noteworthy that the VSV technique with a large kernel and density correction produced results comparable to MC within all VOIs except for the lungs, including the bone. Research Support: R01CA240706 awarded by NCI