Quality Assurance Considerations in Radiopharmaceutical Therapy Dosimetry Using PLANETDose: An International Atomic Energy Agency Study

Implementation of radiopharmaceutical therapy dosimetry varies depending on the clinical application, dosimetry protocol, software, and ultimately the operator. Assessing clinical dosimetry accuracy and precision is therefore a challenging task. This work emphasizes some pitfalls encountered during a structured analysis, performed on a single-patient dataset consisting of SPECT/CT images by various participants using a standard protocol and clinically approved commercial software. Methods: The clinical dataset consisted of the dosimetric study of a patient administered with [177Lu]Lu-DOTATATE at Tygerberg Hospital, South Africa, as a part of International Atomic Energy Agency-coordinated research project E23005. SPECT/CT images were acquired at 5 time points postinjection. Patient and calibration images were reconstructed on a workstation, and a calibration factor of 122.6 Bq/count was derived independently and provided to the participants. A standard dosimetric protocol was defined, and PLANETDose (version 3.1.1) software was installed at 9 centers to perform the dosimetry of 3 treatment cycles. The protocol included rigid image registration, segmentation (semimanual for organs, activity threshold for tumors), and dose voxel kernel convolution of activity followed by absorbed dose (AD) rate integration to obtain the ADs. Iterations of the protocol were performed by participants individually and within collective training, the results of which were analyzed for dosimetric variability, as well as for quality assurance and error analysis. Intermediary checkpoints were developed to understand possible sources of variation and to differentiate user error from legitimate user variability. Results: Initial dosimetric results for organs (liver and kidneys) and lesions showed considerable interoperator variability. Not only was the generation of intermediate checkpoints such as total counts, volumes, and activity required, but also activity-to-count ratio, activity concentration, and AD rate-to-activity concentration ratio to determine the source of variability. Conclusion: When the same patient dataset was analyzed using the same dosimetry procedure and software, significant disparities were observed in the results despite multiple sessions of training and feedback. Variations due to human error could be minimized or avoided by performing intensive training sessions, establishing intermediate checkpoints, conducting sanity checks, and cross-validating results across physicists or with standardized datasets. This finding promotes the development of quality assurance in clinical dosimetry.

Voxel-wise supervised analysis of tumors with multimodal engineered features to highlight interpretable biological patterns

Background

Translation of predictive and prognostic image‐based learning models to clinical applications is challenging due in part to their lack of interpretability. Some deep‐learning‐based methods provide information about the regions driving the model output. Yet, due to the high‐level abstraction of deep features, these methods do not completely solve the interpretation challenge. In addition, low sample size cohorts can lead to instabilities and suboptimal convergence for models involving a large number of parameters such as convolutional neural networks.

Purpose

Here, we propose a method for designing radiomic models that combines the interpretability of handcrafted radiomics with a sub‐regional analysis.

Materials and methods

Our approach relies on voxel‐wise engineered radiomic features with average global aggregation and logistic regression. The method is illustrated using a small dataset of 51 soft tissue sarcoma (STS) patients where the task is to predict the risk of lung metastasis occurrence during the follow‐up period.

Results

Using positron emission tomography/computed tomography and two magnetic resonance imaging sequences separately to build two radiomic models, we show that our approach produces quantitative maps that highlight the signal that contributes to the decision within the tumor region of interest. In our STS example, the analysis of these maps identified two biological patterns that are consistent with STS grading systems and knowledge: necrosis development and glucose metabolism of the tumor.

Conclusions

We demonstrate how that method makes it possible to spatially and quantitatively interpret radiomic models amenable to sub‐regions identification and biological interpretation for patient stratification.

Analysis of differences between 99mTc-MAA SPECT- and 90Y-microsphere PET-based dosimetry for hepatocellular carcinoma selective internal radiation therapy

Background

The aim of this study was to compare predictive and post-treatment dosimetry and analyze the differences, investigating factors related to activity preparation and delivery, imaging modality used, and interventional radiology.

Methods

Twenty-three HCC patients treated by selective internal radiation therapy with ⁹⁰Y glass microspheres were included in this study. Predictive and post-treatment dosimetry were calculated at the voxel level based on ^99mTc-MAA SPECT/CT and ⁹⁰Y-microsphere PET/CT respectively. Dose distribution was analyzed through mean dose, metrics extracted from dose-volume histograms, and Dice similarity coefficients applied on isodoses. Reproducibility of the radiological gesture and its influence on dose deviation was evaluated.

Results

⁹⁰Y delivered activity was lower than expected in 67% (16/24) of the cases mainly due to the residual activity. A mean deviation of − 6 ± 11% was observed between the delivered activity and the ⁹⁰Y PET’s FOV activity. In addition, a substantial difference of − 20 ± 8% was measured on ⁹⁰Y PET images between the activity in the liver and in the whole FOV. After normalization, ^99mTc-MAA SPECT dosimetry was highly correlated and concordant with ⁹⁰Y-microsphere PET dosimetry for all dose metrics evaluated (ρ = 0.87, ρ_c = 0.86, P = 3.10⁻⁸ and ρ = 0.91, ρ_c = 0.90, P = 7.10⁻¹⁰ for tumor and normal liver mean dose respectively for example). Besides, mean tumor dose deviation was lower when the catheter position was identical than when it differed (16 Gy vs. 37 Gy, P = 0.007). Concordance between predictive and post-treatment dosimetry, evaluated with Dice similarity coefficients applied on isodoses, significantly correlated with the distance of the catheter position from artery bifurcation (P = 0.04, 0.0004, and 0.05, for 50 Gy, 100 Gy, and 150 Gy isodoses respectively).

Conclusions

Discrepancies between planned activity and activity measured on ⁹⁰Y PET images were observed and seemed to be mainly related to clinical hazards and equipment issues. Predictive vs. post-treatment comparison of relative dose distributions between tumor and normal liver showed a good correlation and no significant difference highlighting the predictive value of ^99mTc MAA SPECT/CT-based dosimetry. Besides, the reproducibility of catheter tip position appears critical in the agreement between predictive and actual dose distribution.

Predictive value of initial FDG-PET features for treatment response and survival in esophageal cancer patients treated with chemo-radiation therapy using a random forest classifier

Purpose

In oncology, texture features extracted from positron emission tomography with 18-fluorodeoxyglucose images (FDG-PET) are of increasing interest for predictive and prognostic studies, leading to several tens of features per tumor. To select the best features, the use of a random forest (RF) classifier was investigated.

Methods

Sixty-five patients with an esophageal cancer treated with a combined chemo-radiation therapy were retrospectively included. All patients underwent a pretreatment whole-body FDG-PET. The patients were followed for 3 years after the end of the treatment. The response assessment was performed 1 month after the end of the therapy. Patients were classified as complete responders and non-complete responders. Sixty-one features were extracted from medical records and PET images. First, Spearman’s analysis was performed to eliminate correlated features. Then, the best predictive and prognostic subsets of features were selected using a RF algorithm. These results were compared to those obtained by a Mann-Whitney U test (predictive study) and a univariate Kaplan-Meier analysis (prognostic study).

Results

Among the 61 initial features, 28 were not correlated. From these 28 features, the best subset of complementary features found using the RF classifier to predict response was composed of 2 features: metabolic tumor volume (MTV) and homogeneity from the co-occurrence matrix. The corresponding predictive value (AUC = 0.836 ± 0.105, Se = 82 ± 9%, Sp = 91 ± 12%) was higher than the best predictive results found using the Mann-Whitney test: busyness from the gray level difference matrix (P < 0.0001, AUC = 0.810, Se = 66%, Sp = 88%). The best prognostic subset found using RF was composed of 3 features: MTV and 2 clinical features (WHO status and nutritional risk index) (AUC = 0.822 ± 0.059, Se = 79 ± 9%, Sp = 95 ± 6%), while no feature was significantly prognostic according to the Kaplan-Meier analysis.

Conclusions

The RF classifier can improve predictive and prognostic values compared to the Mann-Whitney U test and the univariate Kaplan-Meier survival analysis when applied to several tens of features in a limited patient database.

Monte-Carlo simulations of clinically realistic respiratory gated (18)F-FDG PET: application to lesion detectability and volume measurements

In PET/CT thoracic imaging, respiratory motion reduces image quality. A solution consists in performing respiratory gated PET acquisitions. The aim of this study was to generate clinically realistic Monte-Carlo respiratory PET data, obtained using the 4D-NCAT numerical phantom and the GATE simulation tool, to assess the impact of respiratory motion and respiratory-motion compensation in PET on lesion detection and volume measurement. To obtain reconstructed images as close as possible to those obtained in clinical conditions, a particular attention was paid to apply to the simulated data the same correction and reconstruction processes as those applied to real clinical data. The simulations required 140,000h (CPU) generating 1.5 To of data (98 respiratory gated and 49 ungated scans). Calibration phantom and patient reconstructed images from the simulated data were visually and quantitatively very similar to those obtained in clinical studies. The lesion detectability was higher when the better trade-off between lesion movement limitation (compared to ungated acquisitions) and image statistic preservation is considered (respiratory cycle sampling in 3 frames). We then compared the lesion volumes measured on conventional PET acquisitions versus respiratory gated acquisitions, using an automatic segmentation method and a 40%-threshold approach. A time consuming initial manual exclusion of noisy structures needed with the 40%-threshold was not necessary when the automatic method was used. The lesion detectability along with the accuracy of tumor volume estimates was largely improved with the gated compared to ungated PET images.

Development of a generic thresholding algorithm for the delineation of 18FDG-PET-positive tissue: application to the comparison of three thresholding models

An iterative generic algorithm has been developed to compare three thresholding models used to delineate gross tumour volume on (18)F-FDG PET images. 3D volume was extracted and characteristic parameters were measured. Three fitting models using different parameters were studied: model 1 (volume, contrast), model 2 (contrast) and model 3 (SUV). The calibration was performed using a cylindrical phantom filled with hot spheres. To validate the models, two other phantoms were used. The calibration procedure showed a better fitting model for model 1 (R(2) from 0.94 to 1.00) than for model 3 (0.95) and model 2 (0.69). The validation study shows that model 3 yielded large volume measurement errors. Models 1 and 2 gave close results with no significant differences. Model 2 was preferred because it presents less error dispersion and needs fewer characteristic parameters, making it easier to implement. Our results show the importance of developing a generic algorithm to compare the performances of fitting models objectively and to validate results on other phantoms than the ones used during the calibration process to avoid methodological biases.