Probabilistic clinical target definition with nearest neighbor correlation.

Area of Science:

• Radiotherapy
• Medical Imaging
• Computational Biology

Background:

• Clinical target volume (CTV) delineation in radiotherapy is limited by the invisibility of microscopic disease on medical images.
• Current guidelines propose a probabilistic interpretation of CTV but lack methods to compute microscopic tumor presence (MTP) probability.
• This study addresses the gap in probabilistic MTP estimation.

Purpose of the Study:

• To develop novel stochastic models for estimating the probability of microscopic tumor presence (MTP) at the voxel level.
• To incorporate local spatial correlations within voxel neighborhoods to improve MTP estimation.
• To provide a statistically consistent framework for probabilistic CTV definition.

Main Methods:

• Developed two first-principles stochastic models: constant marginal probability (CMP) and variable marginal probability (VMP).
• The CMP model assumes uniform MTP, suitable for tumors without radial dependence from the gross tumor volume (GTV).
• The VMP model incorporates radial dependence, modeling decreasing MTP with distance from the GTV.

Main Results:

• Both CMP and VMP models accurately reproduced MTP presence fractions.
• CMP model estimated a 0.03 marginal probability of MTP in prostate cancer.
• VMP model replicated radial tumor islet distribution in breast and lung cancer with low mean absolute errors.

Conclusions:

• The proposed stochastic models offer a statistically consistent framework for probabilistic CTV delineation.
• These models enhance the understanding of microscopic disease spread by incorporating local voxel correlations.
• The models provide a novel approach to address uncertainties in radiotherapy planning.