The optimal treatment for uveal melanoma, the most common primary malignant eye tumor, depends on tumor thickness. Conventionally tumor thickness is determined with 2D ultrasound, but MRI allows for a full 3D analysis. It is, however, often difficult to determine the maximum tumor thickness due to its complex 3D shape. We propose a fully automatic framework to segment these MR-images to measure the tumor thickness accurately and evaluate it in four patients. The proposed method has a direct impact on the clinical practice, as a more accurate 3D assessment of the tumor dimensions directly influences therapy determination.
Four UM patients were examined on a Philips Achieva 7-Tesla MRI using a custom-built eye-coil. Eye-motion artefacts were minimized by a cued-blinking protocol3. Two scans were analyzed: a 3D inversion recovery turbo gradient echo (TI/TR/TE/tip: 1280 ms/5.4/2.4 ms/7o, shot-interval:3 seconds, turbo factor 92, resolution 0.53×0.55×0.51 mm3, FOV: 40x42x38 mm3, scan time: 3 minutes) and a 3D T2-weighted fat-suppressed spin echo (TR/TE: 2500 ms/194 ms, resolution: 0.6 mm3, scan time: 3 minutes).
Figure 1 illustrates the framework to segment the tumor. The center of the eye was estimated on the T2-weighted image using the fast-radial symmetry transform4. This center was used to build a mesh to detect the outer border of the sclera and the tumor. The sclera mask was used to register the T2- and T1-weighted images using elastix, with normalized mutual information as a similarity metric5,6. Vitreous body (VB) and lens were segmented on the T1-weighted image7. These masks were subsequently subtracted from the sclera mask to segment the tumor. Finally, tumor thickness was determined by calculating the maximum distance between the tumor and sclera. The analysis algorithms were developed using the rapid-prototyping platform MeVisLab (Fraunhofer MeVis,Bremen, Germany).
The proposed automatic method successfully detected the sclera, tumor, lens and VB. In two patients, there was a small error in the sclera segmentation due to signal loss in the anterior part of the eye (patient 3) and motion artefacts (patient 4). However, This did not influence the subsequent steps and, if needed, can be corrected manually.
The benefit of the automatic tumor segmentation and thickness measurement is clearly illustrated in patients 2 and 3. In patient 2, the tumor is located directly on the optic nerve, making it hard to manually define the boundary between tumor and optic nerve. However, the registration of the T1- and T2-weighted images and the assumption of locally smooth sclera allow for a correct determination of the tumor boundaries. The tumor in patient 3 has a complex shape consisting of multiple lobes. For such geometry, the plane and location for the manual measurement is not known a priori. As the automatic method calculates the distance for the complete tumor boundary, it results in a more accurate measurement.
Part of the retina of patient 3 was detached, a common complication of ocular tumors. The fuzzy clustering correctly identified the tumor as the brightest region. As this may not be valid for all cases, the inclusion of a post-gadolinium scan in the analysis may provide a robust differentiation between tumor and retinal detachment.
Conventionally, the tumor thickness is determined towards the tumor base. By extending this base, however, a smaller tumor thickness can potentially be defined, which might enable more conservative therapeutic options (Figure 5).
Applying the proposed 3D segmentation can also improve proton radiation therapy (PT) planning, which currently relies on manual delineation of the tumor margins on 2D ultrasound images. A 3D MR-based model of the tumor and surrounding structures would allow for a more accurate treatment plan, sparing sensitive structures such as lens and optic nerve.