Introduction
Radiopharmaceuticals therapy (RPT) is a contemporary approach to radiation oncology, which aims to deliver the maximally destructive radiation dose via cancer-targeting radiopharmaceutical [
1]. In the last decades, advances in molecular biology and pharmacology have furnished a wide range of radioactive substances targeting receptors in cancer cell [
2]. Compared to traditional external beam radiotherapy (EBRT), RPT delivers radiation dose more extensively to the intended target tissues and has consequently already proven itself to be effective for the treatment of several metastatic or unresectable cancers through systematic and rationalized administration of the radiopharmaceutical [
3].
However, concerns have been raised about the risk of inadequate balance between therapeutic dose and side effects in RPT. The European Council Directive (2013/59 Euratom) mandates that RPT treatments should be planned according to the optimal radiation dose tailored for individual patients, as has long been the case for EBRT [
4]. The essential requirement of RPT treatment planning is to estimate the absorbed dose in advance of therapy [
5]. The recent European Association of Nuclear Medicine (EANM) procedure guidelines for [
177Lu]Lu-PSMA re-emphasized the value of dosimetry and iterated that exposures of target volumes are to be individually planned and verified.
Theranostics is a unique technology wherein the evaluation of therapeutic agent distribution before treatment informs the treatment protocol [
6]. In the context of PSMA RPT, this involves the pre-therapy positron emission tomography (PET) imaging (e.g. [
68Ga]Ga-PSMA) to determine eligibility for [
177Lu]Lu-PSMA treatment [
7]. Furthermore, post-therapy SPECT/CT serves to estimate tracer distribution and enable dosimetry, facilitating the determination of radiation dose for specific entities such as an organ, tumor, or even a single voxel [
8].
Numerous studies have embraced the concept of theranostics to predict dosimetry in RPT. One extensively explored approach involves physiologically based pharmacokinetic (PBPK) models, which elucidate the fundamental principles underlying the uptake process of pharmaceuticals, including radio-labeled ligands for PET imaging and RPT [
9]. For example, individualized PBPK model parameters can be derived by pre-therapy PET/CT activity concentrations, planar scintigraphy, and tumor volumes, allowing for the individualization of [
177Lu]Lu-PSMA-I&T therapy [
10]. Artificial intelligence (AI) in medicine has burgeoned over the past decade, with machine learning (ML) particularly holding promise for pre-therapy prediction of dosimetry [
11]. Nonetheless, both PBPK-based predictions and our previously developed ML approach are limited to organ-level estimations and do not account for intra-organ heterogeneity, which is crucial for assessing organ toxicity during treatment planning. In the context of RPT, localized areas of high absorbed dose within an organ serve as indicators of organ toxicity, rather than the overall organ dose [
12]. Therefore, accurately predicting intra-organ heterogeneity before initiating therapy is essential for balancing treatment benefits and risks, as well as optimizing therapeutic outcomes.
To address this issue, voxel-level dosimetry has been proposed for patient-specific dose assessment in tumors and organs-at-risk (OAR), aiming to determine the absorbed dose for each image voxel [
13,
14]. Convolutional neural networks (CNNs), a form of deep learning (DL), have emerged as a powerful tool for image synthesis [
15] and can be leveraged to predict voxel-wise absorbed dose map from pre-therapy PET imaging.
This study pursues two main objectives: (1) To investigate the voxel-wise correlation between pre-therapy PET and the therapy dosimetry, thereby characterizing the intra-organ heterogeneity; (2) To explore the feasibility of predicting voxel-wise dosimetry before therapy. We propose two strategies for prediction: (1) Organ-dose guided direct projection; (2) Introducing a novel CNN-based framework named 3D RPT DoseGAN to voxel-wise predict dosimetry. Both strategies are designed to bridge the gap between the distributions of pre-therapy PET imaging and post-therapy dose maps. The overarching aim of this work is to enhance treatment planning for RPT by facilitating accurate predictions of voxel-wise dosimetry.
Discussions
This study aimed to characterize the intra-organ heterogeneity in pre-therapy PET and post-therapy dosimetry, and assess the feasibility of voxel-wise dosimetry prediction. Although prior research has demonstrated a correlation between pre-therapeutic SUV and absorbed dose at the organ level [
17], this correlation has not been explored at the voxel level. Given the spatially heterogeneous distribution of radiopharmaceuticals, resulting in uneven energy deposition, accurate characterization is essential in treatment planning. Our investigation unveiled a moderate correlation between the distribution of pre-therapy PET and dose map. Simulation results (Fig.
1&
S3) indicated a significant positive correlation (
r = 0.99) between PET SUV and absorbed dose, assuming proportional kinetic parameters for [
177Lu]Lu-PSMA I&T and [
68Ga]Ga-PSMA-11. The suboptimal correlation likely arises from differing kinetics in sub-tissues within each organ, as demonstrated by two correlation clusters (
r = 0.40) representing varying kinetic parameters of these sub-tissues.
The organ-dose guided direct projection requires solely the contours of targeted organs from pre-therapy PET imaging and the corresponding SUV features extracted from them. These processes can now be accurately performed using automated tools [
27]. rendering this approach practical in real-world applications. Nevertheless, this method falls short in capturing the dose distribution owing to intra-organ heterogeneity. In contrast, our proposed 3D RPT DoseGAN demonstrates superior prediction accuracy and effectively unveils this heterogeneity. This improvement can be attributed to the implicit extraction of tissue-specific kinetics by deep neural networks from pre-therapeutic PET, enabling the estimation of kinetics for therapeutic dosimetry. By bridging the gap in intra-organ theranostic heterogeneity, our 3D RPT DoseGAN plays a crucial role in determining the radiobiological effect of the treatment. However, it is essential to validate this hypothesis through further pre-clinical studies on a microscopic scale that may not be discernible with clinical imaging techniques. For instance, as shown in Fig.
3, poorer correlations were observed in the liver, possibly due to a larger variety of pharmacokinetics in the sub-tissues. Additionally, although preliminary retrospective analyses indicate that dosimetry imaging could predict prostate-specific antigen (PSA) response [
28], the prognostic imaging biomarker related to overall survival (OS) has not been fully evaluated [
29]. Therefore, the voxel-wise predicted dose map can better assist the development of such biomarkers.
Our study developed a model with only 48 paired theranostic companions, a sample size that may be suboptimal for robust deep learning development. Despite applying augmentation techniques such as patching, the total number of cycles, which is 48, is still insufficient for the development of image synthesis tasks for DL. Additionally, patch-based inputs are highly correlated, which limits the information available for training. Although we attempted to reduce data correlation through random shuffling, a larger dataset would greatly benefit the model in terms of robustness and accuracy. Moreover, the quality of the current dataset has diminished the advantages of voxel-wise dosimetry, as the spatial resolution and field of view of the dosimetry SPECT/CT are limited. We are currently collecting data from our own center using PET with improved sensitivity and resolution (Siemens Vision Quadra) and two different tracers (68Ga and 18F labeled PSMA). Additionally, dosimetry SPECT/CT data with a larger field of view, covering beyond the abdominal region, are being collected. This will enable the prediction of doses for more organs at risk, such as the salivary glands. Furthermore, with the availability of pre-therapy dynamic PET data, models can be developed to predict series of SPECT. Alternatively, with post-therapy PET data, reinforcement learning techniques can be applied to optimize the developed models.
Another bottleneck in our study is the quality of the ground-truth data used for development. Conventionally, voxel-wise dosimetry is conducted using techniques such as voxel S-value (VSV) [
13], dose point kernel (DPK) [
30], and Monte Carlo (MC) simulation [
31]. VSV is a voxel-level implementation of the Medical Internal Radiation Dose (MIRD) schema, which defines S-values. DPK, applied by Siemens Dosimetry Research Tool, measures the absorbed energy per unit mass in a homogeneous water phantom to define radial absorbed dose [
32]. However, both VSV and DPK are limited by their reliance on homogeneous phantoms. MC simulation is a more accurate personalized dosimetry technique that can be applied to heterogeneous activities and media. The “semi” Monte Carlo (sMC) simulation, applied by Hermes Voxel-dosimetry tool, simulates and tracks particles at the voxel-level to estimate absorbed dose by calculating accumulated activity [
33]. However, MC simulation is computationally demanding, time-consuming, and often cumbersome to set up, despite efforts by Hermes to accelerate its tool using simplified methods [
34]. Alternative methods should be explored in further studies due to the limitations of DPK and sMC simulation in terms of heterogeneity issues or computational time and resource requirements. Furthermore, both MC simulation and DPK rely on SPECT and CT images from multiple time points as input, assuming that patients and organs do not move during PET or SPECT imaging. However, patient motion during imaging is inevitable and can result in artifacts and reduced image quality [
35], which consequently affect the accuracy of the dose map used as our ground truth. Additionally, the procedure of co-registration between pre-therapy PET images and dose map images was limited to rigid registration due to the absence of mature deformable registration tools, which may introduce inaccuracies due to changes in patient weight or soft tissue displacements during breathing [
36]. Future studies should consider using advanced registration tools or other possible solutions to improve the accuracy of the ground truth for model training.
Publisher’s Note
Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.