The New Challenge of Decoding Rectal Cancer Signatures By Non-Invasive Imaging: A Retrospective Radiomics Study

Background Bowel cancer is the 4th most common type of cancer in the UK with 14,000 individuals diagnosed with rectal cancer each year. The management of rectal cancer has significantly changed in the last decade with the advent of individualised treatment tailored to the individual patient's pathology and disease burden. Both local control and overall survival in this group of patients are significantly linked with achieving a complete resection with clear circumferential resection margin (CRM).

Radiology is essential to the patient pathway with Magnetic Resonance Imaging (MRI) used in the pre-operative staging process to evaluate not only whether the CRM is potentially involved (tumour or affected lymph node within 1-2mm from CRM), but also other adverse features such as extra mural vascular invasion (EMVI).

With the increasing use of neo-adjuvant chemoradiation therapy (nCRT) the true heterogeneous nature of rectal cancer has become apparent: Rectal cancer patients with similar initial staging have significantly different responses to treatments. MRI can accurately delineate tumour burden; however, it fails to fully take account for this heterogeneity and it is still lacking in adequately evaluating CRM and lymph nodes status. It is only through imaging that the entirety of a rectal tumour is visualised prior to the commencement of treatment. An objective radiological tool that could accurately identify patients who are likely to achieve a complete response to nCRT followed by surgery would have a significant impact in clinical practice by allowing the selection of ideal candidates for organ-sparing strategies.

Recent advances in image acquisition and image analysis that produce quantitative imaging descriptors could potentially play a significant role in bridging this unmet clinical need. The emerging field of Radiomics, uses quantitative data obtained from digital medical images, to extract uncovered mixed biological information. Radiomics is a technique that utilizes all the information available in an image via image processing software. Imaging studies become more than just pictures to be interpreted, instead the wealth of non-visual information generated by computers is used for greater understanding of disease. Within radiomics the numerical data which forms the basis of the images is extracted from a region of interest and analysed producing what is known as radiomics variables. The variables produced are vast and broadly represent the inter and intra-variability between these numerical values. Through the correlation/comparison of these variables with the pathology, genetics and treatment responses the investigators hypothesize that these imaging features (radiomics variables) capture the heterogeneity of rectal cancer. Identifying distinct phenotypic differences of tumours, which are not possible to depict by standard measurements and may have a predictive power and thus clinical significance across different diseases. This technique has been successfully applied in lung and head and neck cancers showing the translational potential of radiomics into clinical practice.

The radiomics variables that can be extracted are divided into primary, second order/texture features and higher order characteristics. Primary characteristics reflect the variables related to the numerical data when assessed alone and includes mean, kurtosis and skewness, predominantly the histogram based characteristics. These do not account for the inter or intratumour heterogeneity as are not representative of the relationship between the voxels. It is secondary order characteristics that reflect how the individual pixels relate to the each other, measuring the intratumoural variability. These variables include fractional analyses and wavelets. Higher order characteristics identify and extract patterns within the region of interest in this case the rectal tumour. Higher order statistics include fractional dimensions and kaplacian transformations. These are a reflection of entire tumours characteristics and the identification of homogeneity within these higher order statistics within population subsets may reveal further information about the underlying tumours biology.