Automatic gross tumor segmentation of canine head and neck cancer using deep learning and cross-species transfer learning

Background

Radiotherapy (RT) is increasingly being used on dogs with spontaneous head and neck cancer (HNC), which account for a large percentage of veterinary patients treated with RT. Accurate definition of the gross tumor volume (GTV) is a vital part of RT planning, ensuring adequate dose coverage of the tumor while limiting the radiation dose to surrounding tissues. Currently the GTV is contoured manually in medical images, which is a time-consuming and challenging task.

Purpose

The purpose of this study was to evaluate the applicability of deep learning-based automatic segmentation of the GTV in canine patients with HNC.

Materials and methods

Contrast-enhanced computed tomography (CT) images and corresponding manual GTV contours of 36 canine HNC patients and 197 human HNC patients were included. A 3D U-Net convolutional neural network (CNN) was trained to automatically segment the GTV in canine patients using two main approaches: (i) training models from scratch based solely on canine CT images, and (ii) using cross-species transfer learning where models were pretrained on CT images of human patients and then fine-tuned on CT images of canine patients. For the canine patients, automatic segmentations were assessed using the Dice similarity coefficient (Dice), the positive predictive value, the true positive rate, and surface distance metrics, calculated from a four-fold cross-validation strategy where each fold was used as a validation set and test set once in independent model runs.

Results

CNN models trained from scratch on canine data or by using transfer learning obtained mean test set Dice scores of 0.55 and 0.52, respectively, indicating acceptable auto-segmentations, similar to the mean Dice performances reported for CT-based automatic segmentation in human HNC studies. Automatic segmentation of nasal cavity tumors appeared particularly promising, resulting in mean test set Dice scores of 0.69 for both approaches.

Conclusion

In conclusion, deep learning-based automatic segmentation of the GTV using CNN models based on canine data only or a cross-species transfer learning approach shows promise for future application in RT of canine HNC patients.

Dose or 'LET' painting--What is optimal in particle therapy of hypoxic tumors?

BACKGROUND

Dose painting is a concept that may increase the tumor control probability (TCP). In particle therapy of hypoxic tumors, it may also be beneficial to redistribute the linear energy transfer (LET) so that the oxygen effect is minimized; so-called LET painting. The purpose of the present study was to use TCP estimates for comparing dose and LET painting of hypoxic tumors.

MATERIAL AND METHODS

Protons, lithium ions and carbon ions were considered. Tumor images tentatively depicting hypoxia were used as input. Optimal dose prescription maps were obtained by optimizing TCP under dose and/or LET redistribution. TCPs were compared to those resulting from conventional particle therapy with no dose or LET painting. The therapeutic gain at a given iso-effect was calculated. Treatment adaptation during therapy in response to changes in the spatial hypoxia distribution was also considered.

RESULTS

Both dose and LET painting gave higher TCPs compared to conventional particle therapy, irrespective of particle type. The therapeutic gain from LET painting, dose painting and combined dose+ LET painting was 1.09/1.43/1.45, 1.24/1.32/1.37 and 1.16/1.23/1.28 for protons, lithium ions and carbon ions, respectively. The importance of treatment adaptation was less pronounced for particles heavier than protons.

CONCLUSION

Dose painting results in higher TCP than LET painting, in particular for protons. For heavier ions, LET painting may also give an enhanced tumor effect compared to conventional particle therapy. Combined dose+ LET painting may only give a marginally increased effect compared to dose painting only. Adaptive carbon ion dose painting seems to be of less importance.

Functional imaging to monitor vascular and metabolic response in canine head and neck tumors during fractionated radiotherapy

Radiotherapy causes alterations in tumor biology, and non-invasive early assessment of such alterations may become useful for identifying treatment resistant disease. The purpose of the current work is to assess changes in vascular and metabolic features derived from functional imaging of canine head and neck tumors during fractionated radiotherapy. Material and methods. Three dogs with spontaneous head and neck tumors received intensity-modulated radiotherapy (IMRT). Contrast-enhanced cone beam computed tomography (CE-CBCT) at the treatment unit was performed at five treatment fractions. Dynamic (18)FDG-PET (D-PET) was performed prior to the start of radiotherapy, at mid-treatment and at 3-12 weeks after the completion of treatment. Tumor contrast enhancement in the CE-CBCT images was used as a surrogate for tumor vasculature. Vascular and metabolic tumor parameters were further obtained from the D-PET images. Changes in these tumor parameters were assessed, with emphasis on intra-tumoral distributions. Results. For all three patients, metabolic imaging parameters obtained from D-PET decreased from the pre- to the inter-therapy session. Correspondingly, for two of three patients, vascular imaging parameters obtained from both CE-CBCT and D-PET increased. Only one of the tumors showed a clear metabolic response after therapy. No systematic changes in the intra-tumor heterogeneity in the imaging parameters were found. Conclusion. Changes in vascular and metabolic parameters could be detected by the current functional imaging methods. Vascular tumor features from CE-CBCT and D-PET corresponded well. CE-CBCT is a potential method for easy response assessment when the patient is at the treatment unit.

Spatiotemporal analysis of tumor uptake patterns in dynamic (18)FDG-PET and dynamic contrast enhanced CT

BACKGROUND

Molecular and functional imaging techniques such as dynamic positron emission tomography (DPET) and dynamic contrast enhanced computed tomography (DCECT) may provide improved characterization of tumors compared to conventional anatomic imaging. The purpose of the current work was to compare spatiotemporal uptake patterns in DPET and DCECT images.

MATERIALS AND METHODS

A PET/CT protocol comprising DCECT with an iodine based contrast agent and DPET with (18)F-fluorodeoxyglucose was set up. The imaging protocol was used for examination of three dogs with spontaneous tumors of the head and neck at sessions prior to and after fractionated radiotherapy. Software tools were developed for downsampling the DCECT image series to the PET image dimensions, for segmentation of tracer uptake pattern in the tumors and for spatiotemporal correlation analysis of DCECT and DPET images.

RESULTS

DCECT images evaluated one minute post injection qualitatively resembled the DPET images at most imaging sessions. Segmentation by region growing gave similar tumor extensions in DCECT and DPET images, with a median Dice similarity coefficient of 0.81. A relatively high correlation (median 0.85) was found between temporal tumor uptake patterns from DPET and DCECT. The heterogeneity in tumor uptake was not significantly different in the DPET and DCECT images. The median of the spatial correlation was 0.72.

CONCLUSIONS

DCECT and DPET gave similar temporal wash-in characteristics, and the images also showed a relatively high spatial correlation. Hence, if the limited spatial resolution of DPET is considered adequate, a single DPET scan only for assessing both tumor perfusion and metabolic activity may be considered. However, further work on a larger number of cases is needed to verify the correlations observed in the present study.

Adaptive radiotherapy based on contrast enhanced cone beam CT imaging

Cone beam CT (CBCT) imaging has become an integral part of radiation therapy, with images typically used for offline or online patient setup corrections based on bony anatomy co-registration. Ideally, the co-registration should be based on tumor localization. However, soft tissue contrast in CBCT images may be limited. In the present work, contrast enhanced CBCT (CECBCT) images were used for tumor visualization and treatment adaptation. Material and methods. A spontaneous canine maxillary tumor was subjected to repeated cone beam CT imaging during fractionated radiotherapy (10 fractions in total). At five of the treatment fractions, CECBCT images, employing an iodinated contrast agent, were acquired, as well as pre-contrast CBCT images. The tumor was clearly visible in post-contrast minus pre-contrast subtraction images, and these contrast images were used to delineate gross tumor volumes. IMRT dose plans were subsequently generated. Four different strategies were explored: 1) fully adapted planning based on each CECBCT image series, 2) planning based on images acquired at the first treatment fraction and patient repositioning following bony anatomy co-registration, 3) as for 2), but with patient repositioning based on co-registering contrast images, and 4) a strategy with no patient repositioning or treatment adaptation. The equivalent uniform dose (EUD) and tumor control probability (TCP) calculations to estimate treatment outcome for each strategy. Results. Similar translation vectors were found when bony anatomy and contrast enhancement co-registration were compared. Strategy 1 gave EUDs closest to the prescription dose and the highest TCP. Strategies 2 and 3 gave EUDs and TCPs close to that of strategy 1, with strategy 3 being slightly better than strategy 2. Even greater benefits from strategies 1 and 3 are expected with increasing tumor movement or deformation during treatment. The non-adaptive strategy 4 was clearly inferior to all three adaptive strategies. Conclusion. CECBCT may prove useful for adaptive radiotherapy.

DCEMRI monitoring of canine tumors during fractionated radiotherapy

PURPOSE

To monitor the contrast enhancement in spontaneous canine tumors during fractionated radiotherapy by Dynamic Contrast Enhanced Magnetic Resonance Imaging (DCEMRI).

METHODS AND MATERIAL

Six dogs with tumors in the oral or nasal cavity received fractionated conformal radiotherapy with 54 Gy given in 18 fractions. T1 weighted DCE imaging was performed prior to each treatment fraction. The tumor was manually delineated in the MR images following every imaging session, and the time dependence of the Relative Signal Intensity (RSI) in the tumor was extracted voxel by voxel. RSI images at the time of maximum enhancement were generated, in addition to images of the initial slope of the RSI curves. The dependence of the median RSI and median slope in the tumor on the accumulated radiation dose was investigated, and images obtained at different treatment fraction were compared by correlation analysis.

RESULTS

Five of the six tumors regressed during treatment. The dose dependence of the RSI varied between the tumors, with some showing an increase and others a decrease in RSI with dose. This was also the case for the initial slope of the RSI curves. The correlation between images acquired before the first treatment fraction and subsequent fractions was in general low, and decreased significantly with accumulated radiation dose for five of six tumors.

CONCLUSIONS

Large individual variations in the dose response of tumor contrast enhancement were found. Decreasing image correlation resulted both from tumor regression and intratumoral changes in the RSI distribution during treatment. These findings may have consequences for treatment design in biological image-guided radiotherapy.