Human–Artificial Intelligence Symbiotic Reporting for Theranostic Cancer Care

Introduction

A comprehensive review of advances in radioligand theranostics in oncology, published 2024 in Molecular Diagnosis and Therapy, ³ concludes that the goal of refining radiotheranostics should be the improvement of the therapeutic index. Yet, this review of state-of-the-art practice does not address quantitation of individual patient parameters necessary for personalized radioligand therapy and optimization of clinical outcomes.

Current approaches to radiopharmaceutical therapy often follow an inflexible “one-size-fits-all” paradigm where the same administered activity is given each cycle regardless of the individual characteristics of patients, or their tumors. Important determinants of response and toxicity, such as tumor heterogeneity, microenvironment, receptor density and affinity, radiosensitivity, radionuclide pharmacokinetics and dosimetry, radioligand pharmacodynamics, and radiomolecular biology, including genomics and immune system status, can all potentially be measured in respect of individual patients, using specific algorithms of artificial intelligence (AI).

A 2024 review of theranostics and AI reports that AI assists tumor characterization, including image interpretation, tumor segmentation, feature identification, and prediction of high-risk lesions. By uncovering potential factors unseeable to the human eye, such as intrinsic variations in tumor radiosensitivity or molecular profile, AI improves diagnostic processes and offers the prospect of precise and detailed evaluation of an individual’s unique clinical profile. ⁴ In oncology, augmenting human capabilities with those provided by AI leads to actionable insights in imaging and therapeutics. The intersection of multiomic data, combined with medical history, social and behavioral determinants, and environmental knowledge, precisely characterizes the molecular biology of cancer and indicates treatment approaches likely to be most effective for the individual patient.

An AI system learns from human experts and real-world cases by collecting feedback, learning from outcomes at all levels and granularities within the system, and continues to improve over time and experience. The process begins with comprehensive data collection drawn from various data types that are multidimensional in nature, including clinical trials, laboratory-based biomarker analyses, and generative AI approaches. This allows modeling of theranostic digital twins, then permitting virtual personalization to optimize treatment of an individual patient. ⁵

AI systems empower and interact with physicians, providing dialogue, visualization, and collaboration, and can translate previously invisible data and information into actionable insights. These insights are tempered, and complemented by human common sense, empathy, morality, and creativity. ⁶

Despite a current opinion expressed in the American Journal of Roentgenology 2024 that AI in nuclear medicine is “More Hype than Reality Today,” ⁷ progress toward general acceptance of human-centered AI systems with the potential for symbiotic collaboration with nuclear physicians is in rapid evolution. ⁸

Given that such an AI-human revolution in clinical practice will come to fruition, the question posed in the European Journal of Nuclear Medicine and Molecular Imaging May 2024 ⁹ “Patient communication of nuclear imaging results: do nuclear medicine practitioners have a role to play?” would be answered unequivocally in the affirmative. The nuclear physician, courtesy of AI, being in possession of radiomics, genomics, and other data relevant to the patient referred for potential theranostic cancer care, would be uniquely placed to collate, analyze, and interpret the quantitative results of imaging and prognosticate likely outcomes of a personalized prescribed dose of molecular targeted radioligand. The nuclear physician would then be most qualified to communicate this integrated, holistic diagnostic information, and the likelihood of effective response, directly to the patient, and thereby obtain truly informed consent to treatment, taking into account the patient’s wishes. The theranostic nuclear physicians would then be effectively reporting to themselves, thence to the patient, and take direct responsibility for the clinical outcome.

The human–AI interactions involved in the imminent, and profoundly disruptive, paradigm shift in future practice of theranostic cancer care require exploration. On offer is a vision for a new integrated approach to nuclear medicine reporting that truly personalizes provision of safe, effective, precision radioligand therapy.

Plato, in The Republic, 375 BCE, wrote the Allegory of the Cave, which may be applied to current qualitative empirical nuclear medicine practice. “Behold! Human beings living in an underground den…like ourselves they see only their own shadows, or the shadows of one another, which the fire throws on the opposite wall of the cave.” ¹⁰ The prisoners in Plato’s Cave do not know what they do not know; they do not even know that they do not know. They dwell in ignorance but cannot recognize it.

In Plato’s account, the unenlightened must rely upon accident or the beneficent intervention of others for the critical first step; “one of them was freed.” What follows his release is not a swift and purposeful escape motivated by eager anticipation of the waiting outside world: it is only the slow, hesitant, gradual, painful process of learning itself. This is an educational intervention. It is initiated from without, and it is initially coercive, requiring the forceful overcoming of the learner’s resistance. Eventually, as understanding flows into him, he finally comes to know the wondrous sunlit real world. In general, human beings tend to prefer cognitive comfort, the reinforcement of the familiar, to an encounter with the unknown. ¹¹

It is time to emerge into the light of day, guided by AI enlightenment, into a realizable world of optimal, precise, personalized targeted molecular radioligand theranostic care of cancer. “Is it really happening?” Commentary in the Journal of Nuclear Medicine, published September 5, 2024, ¹² remarks that: “The continuing development of molecular radiotherapy stubbornly refuses to take individual patient dosimetry into account…a visionary reorganization of practice is required… [to redress] what effectively constitutes a willful blindness.”

Prostate Cancer Radioligand Prostate-Specific Membrane Antigen Theranostics

An instructional article on “How to Report PSMA PET” in Seminars in Nuclear Medicine in 2023 reviewed current published reporting criteria and emphasized the need for harmonization and synoptic reporting. However, documentation of the several criteria proposed by the European Association of Nuclear Medicine, Prostate Cancer Molecular Imaging Standardized Evaluation (PROMISE), and prostate-specific membrane antigen (PSMA)-RADS shows that each provides different insights in interpretation of PSMA-PET studies and different recommendations on how to draft the diagnostic report. ¹³ Tumor-to-reference organ ratio and SUV_max are, most commonly, used to determine patient candidacy for PSMA radioligand therapy of prostate cancer but may be compromised by reference-organ variability, technical changes in SUV_max, and partial volume effect. Image screening criteria may need to be redefined for different prostate cancer targeted radioligands and are likely to be further complicated by combination therapies such as with cabazitaxel, enzalutamide, novel immunotherapies, and poly-ADP ribose polymerase inhibitors. This comprehensive review of state-of-the-art reporting practice in theranostics concludes that “quantitation of PSMA-PET parameters and AI algorithms will almost certainly revolutionize the way PSMA-PET will be reported in the near future.” ¹³

Corroboration appeared in the same journal in 2023 in a review of PSMA-PET AI applications, which concluded that “the emergence of precision medicine comes with the need for quantitative data, ie imaging biomarkers. The common qualitative way to report PET/CT studies will not be sufficient and manual segmentation of tumors is a very time-consuming task…AI tools will support nuclear medicine physicians to improve interpretation accuracy, decrease interreader variability, save reporting time and provide clinicians with quantitative data.” ¹⁴

Projections for a prostate cancer PET/CT imaging study based upon the Society of Nuclear Medicine and Molecular Imaging AI Task Force evaluation and validation of AI algorithms and radiomic studies have been applied in a chain from radiochemistry to physician report generation. ¹⁵ AI could first predict drug–target interactions, predict and optimize radiochemical reactions, carry out de novo drug design, and optimize radiopharmacy workflows. AI machine learning methods may resolve difficult issues in image acquisition and instrumentation, and for image reconstruction, increase speed and deliver better signal-to-noise ratio and fewer artifacts. Image analysis can be automated using AI for different tasks such as lesion detection, segmentation and quantification for diagnosis, and dosimetry. Moreover, AI has the potential to investigate patterns associated with patient results within large biological and imaging datasets. In addition, AI can also detect and interpret diagnostic images for translation into reports for patients and for clinical data bases. Finally, in concert with nuclear physicians, AI facilitates actionable advice to clinicians, and to the patient, after extracting, distilling, and integrating clinical information from various sources.

Automated Prostate Molecular Imaging Standardized Evaluation (aPROMISE) is a deep learning algorithm capable of high sensitivity in the detection of prostate cancer lesions in ¹⁸F-DCFPyl PSMA PET/CT imaging, with consistency in quantitative assessment across multiple human readers. ¹⁶ This AI application still requires physician review of the image to select lesions for evaluation but saves labor and standardizes the report. It also demonstrated disparity of outcome based on reader experience in PSMA imaging. The aPROMISE software has FDA approval. It first analyses the CT component of the PSMA PET/CT to automatically segment it into anatomical regions, subsequently the PSMA PET image is analyzed to detect metastases and merged with quantified tracer uptake to generate the miPSMA score. ¹⁷ Thus by leveraging deep learning, aPROMISE automates the labor-intensive task of anatomical segmentation and PSMA uptake quantification.

While AI can detect lymph node involvement and metastatic disease with high accuracy (area under the curve of 98%) and sensitivity (62–97%), ¹⁴ can estimate tumor burden for prognostic purposes, and predict tumor response of metastatic prostate cancer, these AI models are only as good as the data upon which they are trained. In the absence of multiple biopsies, the “ground truth” of these AI models is limited by visual diagnosis by experienced nuclear physicians and their human ability to distinguish true-from false-positive lesions. It should also be acknowledged that AI has the potential to detect occult lesions that human readers are unable to see.

The synergy of human–AI interaction to enhance prostate cancer diagnosis and grading is illustrated by the significant improvement in accuracy of Gleason grading of biopsy slides when pathologists are assisted by AI, in comparison with unassisted colleagues, or with AI alone. ¹⁸

The consummate ability of AI to collect and collate Big Data engenders the prospect of an integrated report of a theranostic imaging study to encompass all the relevant clinical data and pathology, genomics, and radiomics. This characterization of molecular biology may be anticipated to allow prediction of response to prescribed dose radioligand therapy, and assessment of prognosis. The nuclear physician writing such a comprehensive holistic report will thus be in a uniquely privileged position to provide optimal personalized efficient theranostic cancer care.

Dosimetry

Notwithstanding the nuclear medicine maxim: We treat what we see, and we see what we treat, prostate cancer theranostics is currently given blindfold with respect to actual radiation absorbed dose (Gy) delivered to the tumor, kidneys, bone marrow, and salivary glands. All patients receive a standard fixed activity (GBq), ¹⁷⁷Lu-PSMA without measurement of biologically effective dose (BED). It is well established that fixed activity administrations to all patients deliver a wide range of radiation absorbed doses to tumors and to tissues-at-risk, which may vary from patient to patient by an order of magnitude, ¹⁹ raising the probability of under- and overtreatment. So-called standardized administered activity (which many erroneously call the “dose”) may result in toxicity with a small volume of disease, while significantly undertreating patients with a large tumor burden. ²⁰ Furthermore, there is a weak scientific foundation for assessment of radiation effects and prediction of probabilities of effectiveness and risks of toxicity based upon administered activities. ²¹

The consequence of this one-size-fits-all approach is that the vast majority of patients are undertreated and the clinical outcome is unsatisfactory. ²² When responders are compared with nonresponders, their tumor radiation absorbed doses are significantly higher in prostate cancer patients treated with ¹⁷⁷Lu-PSMA. Dosimetry performed on the first cycle showed a mean up to 80 Gy deposited in the index tumor of those who responded compared with half that in nonresponders. The difference in mean tumor dose between responders and nonresponders was three times higher (90 Gy vs. 30 Gy). ²³

It is inconceivable that a radiation oncologist would contemplate giving radiotherapy without prescribing the tumor dose in Gy, and subsequently validating the delivery of that planned dose. In contradistinction, prevalent nuclear medicine theranostic clinical practice contravenes the European Council Directive 2013/59/EURATOM, in which it is stated: “For all medical exposure of patients for radiotherapeutic purposes exposures of target volumes shall be individually planned and their delivery appropriately verified.” Specifically, the ICRP report 140 states that: “Individual absorbed dose estimates should be performed for treatment planning and for postadministration verification of doses to tumors and normal tissues.” ²¹ In routine clinical theranostic practice worldwide, the radiation absorbed dose (Gy) to tumor and critical normal organs is neither prescribed on pretherapy quantitative imaging studies, nor measured and validated after each cycle of ¹⁷⁷Lu-PSMA treatment.

In this absence of individualized dose (Gy) prescription and delivery based upon dosimetry, the current status of theranostic care of prostate cancer cannot be characterized as either precision or personalized medicine. The advent of AI-assisted dosimetric methodology renders this empirical practice inexcusable and unacceptable.

Practical ways in which AI may facilitate dosimetry calculations can be enumerated as follows: Registration of multimodality and multitime point pre- and postimaging studies is necessary for subsequent segmentation, treatment planning, image-guided radiotherapy, and response assessment. AI techniques have shown better accuracy and robustness compared with conventional registration methods. They can be more easily generalized across different modalities and can mitigate the effects of image artifacts and noise. ²⁴

Segmentation allows measurement of activity within each organ and tumor as well as the estimate of corresponding mass of each volume of interest. Manual delineation is time-consuming and subject to intraobserver and interobserver variability. Validated AI-based models for fully automated, robust, accurate segmentation of organs and lesions in PET, PET/CT, and SPECT/CT images can help delineate organs-at-risk, and tumors, to achieve a personalized dosimetry framework. ²⁴

AI models can be pretrained on PET/CT images and then “tuned” using SPECT/CT data and achieve feature mapping between domains to facilitate estimation of the probability that a voxel in the SPECT image belongs to a tumor or critical normal organ. AI can use the information from diagnostic scans and therapeutic cycles to improve the assessment of time–activity curves and time-integration of activity to improve the characterization of radiation absorbed dose in subsequent therapy cycles where volumes of interest may be changed by tumor shrinkage or progression.

AI can potentially be used to combine multiscale dosimetry knowledge for accurate, effective dose modeling and help to elucidate the complex relationship between pretherapy patient data such as imaging radiomics, demographic data, genomics, and laboratory results, and the predicted radiation dose distribution to be obtained by a prescribed therapeutic administered activity. These multimodality data could then be incorporated into an integrated nuclear medicine report.

Conventionally, voxel-wise dosimetry is conducted using techniques such as voxel S-value, dose point kernel, and Monte Carlo simulation, but these are limited by their reliance on homogeneous phantoms, which compromise calculation of dosimetry in heterogeneous tissue uptake. Deep learning approaches have the potential to predict voxel-wise heterogeneous dose maps and to perform prospective dosimetry for ¹⁷⁷Lu-PSMA therapy based on pretherapeutic PSMA PET/CT. ^25,26

Standard dosimetric techniques require major physics resources and computational time, whereas AI efficiency in workflow is dramatic. The calculated dose difference of deep convolutional neural network-derived absorbed dose-rate maps against Monte Carlo dosimetry of ⁶⁸Ga-NOTARGD PET/CT was less than 2%, with a time effort of less than 4 min, compared with more than 235 h of computation time of Monte Carlo simulation. ²⁷

There is an untapped potential to apply radiomics analysis to molecular imaging, both from pretherapy and post-therapy images from CT, PET, MR, and SPECT modalities and combine radiomic features detectable from within the 3D absorbed dose maps to better predict individual clinical outcome. A deep network can directly extract and identify the most predictive radiomic features and has the potential to incorporate this information into its own evolving capability to enhance performance.

As of 2023, there were 4 software products approved by the FDA for radiopharmaceutical therapy dosimetry, but only for retrospective post-therapy use. ²² Regulatory issues also constrain clinical application of a prescribed, personalized, optimal radiation absorbed dose, since it is not possible under the present U.K., European, or North American product licenses to increase administered activity, and thus possibly enhance tumor kill, even if dosimetry shows that this can be done without exceeding the maximum tolerated radiation absorbed dose. ²⁸ It is difficult to imagine radiation oncologists tolerating such bureaucratic regulatory interference in their clinical practice, with its inevitable compromise of the efficiency of cancer treatment of their individual patient.

Paradoxically, patient-specific dose planning in external beam radiotherapy (EBRT) has impaired the effectiveness of theranostic cancer treatment in which the EBRT dose limit of 23 Gy to kidney has been arbitrarily, and erroneously, applied to radioligand therapy. ²⁹ This conservative limit, if applied to ¹⁷⁷Lu-DOTATATE peptide receptor radionuclide therapy, results in 80% of patients with neuroendocrine tumor (NET) being undertreated. In fact, a cumulative renal BED of 40 ± 2 Gy in patients with NET treated with ¹⁷⁷Lu-DOTATATE experienced no significant renal toxicities. ³⁰

The proposition that the level of activity administered (GBq) empirically has a greater impact on treatment outcome than the subsequent biodistribution, the radiation delivery and the absorbed dose (Gy) are ignoring the results of decades of radiation research on biological systems. ³¹

Pretherapeutic tracer dosimetry of ¹⁷⁷Lu-PSMA in prostate cancer patients is now practical using standard SPECT/CT imaging equipment and commercially available dosimetric software to estimate tumor radiation absorbed dose, and dose to kidneys and bone marrow enabling prospective prescription of effective, nontoxic therapeutic administered activity to an individual patient. ^32,33 Validation of this predictive dosimetry, using quantitative post-therapy SPECT/CT imaging, confirmed the accuracy of pretherapy tracer dosimetry to prospectively determine radiation absorbed dose delivery in tumors and in critical normal organs in each patient.

Prediction of hematological and bone marrow toxicity is essential, and will be facilitated by deep learning algorithms, which can be effectively deployed in low-activity tracer dose administrations. AI can facilitate the linkage between macrodosimetry and microdosimetry, ultimately enabling the generation of a comprehensive patient profile. Thus, AI-assisted predictive dosimetry, combined with clinical parameters, will enable personalized theranostics and, in particular, integration of AI methodologies in bone marrow dosimetry will help to preclude hematological toxicities. ³⁴

Elements of AI–Nuclear Physician Relationship

Render therefore unto Caesar the things which are Caesar’s: and unto God the things that are God’s. Matthew 22:21

AI computer algorithms have no soul, no inherent morality, nor ethical obligation, and are incapable of empathic, compassionate emotional engagement with human beings. AI is not a moral agent and not a human surrogate, it cannot be held responsible, nor can it be accountable for its own actions. It is important to recognize at the outset that the commercial AI tendency is to deploy plagiarized affective sophistries programmed to simulate a semblance of quasiemotional relationships with human beings. AI is potentially more intelligent and imaginative than us, which one may find difficult to understand or control. It is imperative that the nuclear physician seeking a symbiotic relationship with AI, to optimize personalized precision theranostic cancer care, does not enter into a Faustian pact. The components of the interaction must be clearly delineated and be subject to continuing and assiduous personal physician oversight and control. Explainable AI algorithms are under development with the objective of achieving a human–AI interface, which interactively integrates the existing a priori knowledge, experience, and conceptual understanding of human experts into statistical learning methods. ³⁵ The aim is not to replace the human decision-makers with AI, rather it is to produce accurate algorithmic predictions, which can then be supplemented with the value judgments, creativity, intuition, emotions, and feelings of empathic human physicians in a phronetic relationship with their patient.

Today’s machine learning, relying on statistical learning algorithms, large datasets, and available computational capacity, should, in principle, enable evidence-based decision-making across various domains by replicating statistical frequencies from previous data and improving them based on new data. However, such machine learning faces challenges in explainability, including sense-making, consideration of context, and decision-making under uncertainty, making it necessary to incorporate human expertise for usable intelligence. Overall, AI alters the decision-making process, and therefore, humans must learn how to think differently in and about decision-making to benefit from using AI. Physicians will require a new mental model, comprising both AI predictions and human judgments, where humans and AI must learn to adapt to each other. ³⁶

Issues such as explainability, agency, and accountability are central to human–AI interaction. The ability to understand and predict what an AI-based algorithm is doing has important consequences for physician’s trust with respect to the computational models. ³⁷

Trust is a key dimension for acceptability of AI by the nuclear medicine professional, and the patient. ³⁸ Genuine trust is a human-to-human interpersonal concept that depends on rich affective and normative attitudes. Applying trust to AI is a category error mistakenly assuming that AI belongs to a category of things that can be trusted. ³⁹ AI systems are not the appropriate objects of trust under any familiar philosophic accounts of trust. AI systems can be relied on, and are capable of reliability, but cannot be trusted. “I rely on you when I predict that you will behave in a certain way, though I trust you when I judge that you ought to behave in a certain way.” ⁴⁰ Unlike a human clinician, AI systems have no goodwill toward us, nor any motivation to act in our interests. Thus, ascribing trustworthiness to AI would appear to indicate conceptual misunderstanding, and human interaction and mediation in AI medical applications should be deemed essential to engender trust of patients. ⁴¹

The Society of Nuclear Medicine and Molecular Imaging Strategic plan 2023 emphasizes the importance of human agency and the necessity for AI systems to empower physicians and patients allowing them to make better informed decisions and to foster their autonomy. ⁴² A new set of skills, including physician oversight and interaction with AI tools, will evolve and must be refined.

Fundamentally, patients want to have a basic understanding of the AI technology, how it will be applied to their own care, know that the application of this technology is safe, be ensured that the application results in increased efficiency and quality of care, know that this application ultimately enables deeper personal interactions with their physician, and trust that both the provider and developers of the technology are accountable for the outcome. ⁴³

Other ethical questions that arise when using human data to develop human-targeted AI applications include informed consent, privacy, data protection, data ownership, objectivity, and inequity, ⁴⁴ and they remain to be adequately addressed.

Hitherto, AI has been essentially unregulated. UNESCO published recommendations on the ethics of AI in 2020 and these are enumerated in a 2023 exposition of AI principles, ethics, and key requirements for responsible regulated AI systems. ^45,46 The recently promulgated EU AI Act 2024 will regulate AI in the European Union by 2026, and entered into force across all EU member states August 1, 2024. ⁴⁷ Proposed mandatory regulatory guardrails for AI are under current governmental deliberation in Australia. ⁴⁸ They incorporate the following principles:

Establish, implement, and publish an accountability process, including governance, internal capability, and a strategy for regulatory compliance.

The adoption of formal government regulation of AI should help reassure physicians seeking to use AI algorithms to enhance their practice of theranostics. The challenge for the physician is to develop a new mindset to manage the complexities of AI algorithms, understand the strengths and limitations of the numerical AI analyses, master the ethical translation to empathic patient care, and be responsible and accountable for clinical decisions and outcomes based on AI.

Paradoxically, AI-physician symbiotic theranostic cancer care has the potential to enable deeper, more meaningful, interactions between the patient and the nuclear physician by liberating the cognitive and emotional space otherwise taken up by routine manual image analysis and exigencies of reporting, and attending to bureaucratic demands of the electronic medical record.

Conclusion

Human–AI symbiotic reporting will allow the nuclear physician to generate, and act upon, a comprehensive, integrated compilation of personalized radiomic, dosimetric, genomic, and clinical and laboratory data. This holistic report would facilitate selection of appropriate patients and then permit prescription of an optimized administered activity prospectively calculated to deliver a preordained individualized precise radiation absorbed dose to tumor, designed to spare critical normal organs. It would also predict the probability of tumor response and a likely clinical outcome.

Ultimately, the comprehensive AI-assisted integrated nuclear medicine report will be directed to the nuclear physician, given that it will become apparent that the theranostic clinician is uniquely qualified to explain the proposed treatment to that individual patient, foresee the likely outcome, obtain truly informed consent, and personally administer the prescribed theranostic dose of radioligand.

Continuing collaboration of AI can then validate, by post-therapy imaging, the actual radiation absorbed dose delivered to target tumor, and to organs-at-risk. AI radiomic analysis of quantitative post-treatment scans could then be used to prescribe a tailored dose for subsequent therapy cycles, predicated upon tumor response and measured pharmacodynamics, estimated radioresistance, and determination of other molecular biological variables capable of being characterized by AI methodology.

Clinical implementation of AI algorithmic assistance in theranostic practice will require extensive physician training, and a new mindset, since it is the nuclear physician who will take personal responsibility, and be accountable, for patient outcomes. The trusting, empathic relationship is, as always, forged between the doctor and patient and is not able to be delegated to AI. Acceptance of this challenge will allow, for the first time, truly personalized precision radiomolecular therapy of cancer.