CHAPTER 17 Automatic Tumor Growth Detection
INTRODUCTION
Most meningiomas are slowly growing lesions that can occur on the surface of the brain, the skull base, the dural reflections, or within the ventricles. About 90% of these tumors are classified as benign.1 Neurosurgeons generally avoid operating on patients with small meningiomas, particularly those that are difficult to access surgically, and prefer to carefully monitor tumor growth.2 Monitoring includes neurologic evaluation and periodically acquiring magnetic resonance scans of the patient’s brain. Neuroradiologists and clinicians visually inspect the scans for evidence of tumor growth. In clinical practice, finding evidence for subtle growth can be very difficult, particularly between scans taken at relatively short intervals. The main reason for the difficulty is that any changes in the head position or variations in the intensity profile between scans can actually obscure the slow growth that is so characteristic of meningiomas. In addition, very small changes in the linear dimensions seen on cross-sectional imaging can reflect appreciable volumetric change.
COMMON PROCEDURES FOR MEASURING TUMOR GROWTH
Surgeons and oncologists frequently analyze the evolution of tumors by viewing brain scans via a light box or visualization software. The analysis often includes sophisticated measuring techniques, as simple visual inspections of the images typically ignore small tumor growth (Fig. 17-1). These techniques commonly measure the size of the tumor individually for each scan, then compute the growth by combining the measurements of consecutive scans. Because change in the tumor dimension may be so difficult to appreciate between any two sequential scans it becomes critical that the clinician compare the most recent scan to the earliest available image as only by comparing these may the growth of the tumor be visibly evident (see Fig. 17-1 for an example). Thus, the patient may undergo extra testing and tumor growth may not be promptly detected. In addition, such methods do not provide a quantitative measure that could give an indication of the rate of growth and thus aid in treatment decisions. One metric for performing this task is the World Health Organization (WHO) response criteria.3 The criteria infer the size of each tumor through bidimensional measurements, which are the tumor’s largest diameter and perpendicular diameter.4 To increase efficiency and reproducibility, the Response Evaluation Criteria in Solid Tumors (RECIST)5 is based only on the largest diameter. However, both measurements ignore small growth deviating from the largest diameter directions.6,7 More accurate measures have come from manual segmentations of the tumor volume, but these approaches are very labor intensive, and are sensitive to large variations in experts identifying tumor regions.8 To streamline the process, researchers have developed automatic methods.9,10 These methods generally outline pathology by combining the image data with general information about the visual appearance of healthy tissue and pathology. The automatic methods are, however, still impacted by intrarater variability caused by image artifacts or altered head position, as they determine the tumor volume separately for each scan. To date, there has not been a technique that has been widely adopted by clinicians.
In recent years, a new line of automatic tools has emerged that analyzes growth of lesions by processing a sequence of scans simultaneously.11–13 These tools first normalize the sequence of images and then relate unusual patterns across the scans to regions of change. In this spirit, Rey and colleagues14 aligned the series of scans of multiple sclerosis patients to each other, a process that resulted in a deformation map that defined region-specific transformations from one scan to the next. They then showed that growing lesions produce patterns in the maps that are inherently different from dormant tissue. Angelini and colleagues15 proposed an alternative approach for brain tumors in which expected variations across the scan due to magnetic resonance acquisition were accounted for by aligning the head in each scan to a fixed pose and equalizing the global intensity patterns across the scans. They then related gross regional differences in intensity patterns across the scans to tumor growth.
To the best of our knowledge, state-of-the-art software tools are tested on scans with visible tumor/lesion growth only, as was the case for Rey14 and Angelini and colleagues.15
