Cohort-based T-SSIM Visual Computing for Radiation Therapy Prediction and Exploration

Abstract

We describe a visual computing approach to radiation therapy (RT) planning, based on spatial similarity within a patient cohort. In radiotherapy for head and neck cancer treatment, dosage to organs at risk surrounding a tumor is a large cause of treatment toxicity. Along with the availability of patient repositories, this situation has lead to clinician interest in understanding and predicting RT outcomes based on previously treated similar patients. To enable this type of analysis, we introduce a novel topology-based spatial similarity measure, T-SSIM, and a predictive algorithm based on this similarity measure. We couple the algorithm with a visual steering interface that intertwines visual encodings for the spatial data and statistical results, including a novel parallel-marker encoding that is spatially aware. We report quantitative results on a cohort of 165 patients, as well as a qualitative evaluation with domain experts in radiation oncology, data management, biostatistics, and medical imaging, who are collaborating remotely.

Fig. 1.

Visual computing for cohort-based radiation therapy prediction. A stylized view of the predicted radiation plan of the current patient is placed centrally; top pale markers (front and back of eyes) receive the least radiation; tumors in black receive the most. Additional RT views show the most similar patients under our novel T-SSIM metric, who contribute to the prediction; the most similar patient is currently selected for comparison. A scatterplot (left) shows 4 clusters generated through the T-SSIM metric, with the current (cross) and comparison patient highlighted. A parallel-marker encoding (bottom) shows the predicted (blue cross) per-organ dose distribution within the context of the most similar patients; spatially collocated organs are in contiguous sections of the x-axis.

Fig. 2.

Construction of the spatial similarity measure. (A) A sliding window (a sphere, illustrated in 2D here) steps through the centroids of the organs to identify nearby organs. (B) Each step in the sliding window is used to constructed a variable-length vector using the set of nearby organs (e.g., 2 organs in Step 1, 3 in Step 2, 4 in the n step). (C) Two sets of vectors populated with tumor-organ distances and volumes, respectively, for each patient are created. These vectors are used as inputs into a similarity function (T-SSIM) to compare two patients. The vectors can be represented in matrix form, as described in Subsection 3.4.

Fig. 3.

Three stylized renderings of the 3D radiation plan for Patient 152 showing the actual (left), the predicted (center), and the prediction error (right, in blue) in the radiation plan. Circular markers indicate the location of organs at risk, and black markers indicate the tumors. Red luminance is mapped to the radiation dose (higher dose mapped to darker shades) and blue luminance is mapped to error size, respectively. Transparent organ models are shown for context. The pale markers at the top correspond to the eyes, and the lowest marker is located down the spine.

Fig. 4.

Two configurations of the scatterplot. The data can be plotted across the principal components of the radiation doses (top), primary and secondary tumor volumes (bottom), and principal components of the distances between each organ and the primary tumor volume (see Fig. 1) top left.

Fig. 5.

Example radiation plans for the 4 different patterns identified in the data. Top left: a plan with a higher dose to the lower-anterior throat. Top right: a plan with a ‘standard’ dose distribution, where radiation is lower in the throat and distributed to both the left and right sides of the head. Bottom right: a plan with dosing primarily to the right side of the head. Bottom left: a plan with dosing primarily to the left side of the head.

Fig. 6.

Snapshots of key moments during the qualitative evaluation. (A) Picture of the dose-PCA scatterplot on the reduced cohort using the clustering provided by GC. Clusters visibly divide the feature space despite being done without dose information. (B) RT plan for the patient being inspected (shown in (A) as the cross magenta marker). (C) RT prediction error for the patient. Error rates are highest on the left side of the head. (D) Close up of the dose-distribution. One of the matches (highlighted) is significantly further from the other matches. (E) Parallel-marker dose plot of the patient and its matches. Doses from the suspicious match (highlighted) are significantly lower for several adjacent areas. (F) Radiation plan of the suspect patient, who received almost no radiation to the left side of their head.