2949 - A Rule-Based AI Expert System for Assisted Quantitative Small Target Oropharyngeal PET 18F-FMISO Hypoxia Imaging

Purpose/Objective(s):

Oropharyngeal tumor 18F-FMISO PET hypoxia assessment is a valuable tool for personalized chemoradiation therapy, but interpretation is highly subjective due to well-known partial volume effects (PVE) underestimating the primary quantitative parameter SUVmax in small targets. This work presents a transparent rule-based AI expert system to objectively correct for PVE, mitigating interpretive variations and enabling widespread reliable implementation of this procedure for new users.

Materials/Methods:

The system utilizes two sequential inference engines. The first inference engine (IE1) determines a patient-specific Point Spread Function (PSF) to account for clinical variations in system resolution. This is accomplished by measuring the mean penumbra width (10%-90%) of a relatively large internal reference PET avid structure in the patient's PET scan near the target of interest. The measured penumbra is then associated, using a pre-calculated theoretical model, to the standard deviation (SD) of the Gaussian PSF that would produce the same edge response in an ideal step function.

The second inference engine (IE2) determines the PVE correction by querying a multi-dimensional lookup table. The table was generated by computationally modeling a parameter space consisting of ideal edge functions of varying Full Widths at Half Maximum (FWHMs) convolved with a range of Gaussian PSFs. For each resulting blurred profile, the source intensity correction factor (K) required to produce the post-convolution measured SUVmax was calculated and stored. For clinical use, first the patient's mean target profile FWHM is measured. The patient-specific PSF determined by IE1 and the measured target FWHM serve as inputs to the lookup table. The correction factor K as a function of the two measurable parameters times the measured SUVmax then represents the true intensity of the target without partial volume effects

Results:

For a representative case with a 16.5 mm penumbra width reference structure, the system derived a patient-specific PSF FWHM equal to 2.355*SD = 6.0 mm. For a 10.0 mm FWHM target within the same patient, the system's lookup table determined that the true, corrected SUVmax is 1.152 times the measured value. This correction is clinically significant considering that the standard quantitative hypoxia-positive criteria is target SUVmax to reference ratio above 1.2. Therefore, a target with a measured ratio of only 1.04 can also be correctly reclassified as positive after correction since1.04 × 1.152 ˜ 1.2.

Conclusion:

The rule-based AI expert system accurately quantifies SUVmax by correcting for partial volume effects in small tumors using patient-specific data. By providing an objective metric, this tool reduces subjectivity and may explain the historical reliance on visual overrides based on years of experience. This work provides a robust, transparent method to standardize 18F-FMISO PET quantitative imaging used in patient selection for personalized therapies.