Defining the loss function is an important part of neural network design and critically determines the success of deep learning modeling. A significant shortcoming of the conventional loss functions is that they weight all regions in the input image volume equally, despite the fact that the system is known to be heterogeneous (i.e., some regions can achieve high prediction performance more easily than others). Here, we introduce a region-specific loss to lift the implicit assumption of homogeneous weighting for better learning. We divide the entire volume into multiple sub-regions, each with an individualized loss constructed for optimal local performance. Effectively, this scheme imposes higher weightings on the sub-regions that are more difficult to segment, and vice versa. Furthermore, the regional false positive and false negative errors are computed for each input image during a training step and the regional penalty is adjusted accordingly to enhance the overall accuracy of the prediction. Using different public and in-house medical image datasets, we demonstrate that the proposed regionally adaptive loss paradigm outperforms conventional methods in the multi-organ segmentations, without any modification to the neural network architecture or additional data preparation.

To minimize the adverse dosimetric effect caused by tumor motion, it is desirable to have real-time knowledge of the tumor position throughout the beam delivery process. A promising technique to realize the real-time image guided scheme in external beam radiation therapy is through the combined use of MV and onboard kV beam imaging. The success of this MV-kV triangulation approach for fixed-gantry radiation therapy has been demonstrated. With the increasing acceptance of modern arc radiotherapy in the clinics, a timely and clinically important question is whether the image guidance strategy can be extended to arc therapy to provide the urgently needed real-time tumor motion information. While conceptually feasible, there are a number of theoretical and practical issues specific to the arc delivery that need to be resolved before clinical implementation. The purpose of this work is to establish a robust procedure of system calibration for combined MV and kV imaging for internal marker tracking during arc delivery and to demonstrate the feasibility and accuracy of the technique. A commercially available LINAC equipped with an onboard kV imager and electronic portal imaging device (EPID) was used for the study. A custom built phantom with multiple ball bearings was used to calibrate the stereoscopic MV-kV imaging system to provide the transformation parameters from imaging pixels to 3D world coordinates. The accuracy of the fiducial tracking system was examined using a 4D motion phantom capable of moving in accordance with a pre-programmed trajectory. Overall, spatial accuracy of MV-kV fiducial tracking during the arc delivery process for normal adult breathing amplitude and period was found to be better than 1 mm. For fast motion, the results depended on the imaging frame rates. The RMS error ranged from approximately 0.5 mm for the normal adult breathing pattern to approximately 1.5 mm for more extreme cases with a low imaging frame rate of 3.4 Hz. In general, highly accurate real-time tracking of implanted markers using hybrid MV-kV imaging is achievable and the technique should be useful to improve the beam targeting accuracy of arc therapy.

By analyzing backscattered echo signal power spectra and thereby obtaining backscatter coefficient vs. frequency data, the size of subresolution scatterers contributing to echo signals can be estimated. Here we investigate trade-offs in data acquisition and processing parameters for reference phantom-based backscatter and scatterer size estimations. RF echo data from a tissue-mimicking test phantom were acquired using a clinical scanner equipped with linear array transducers. One array has a nominal frequency bandwidth of 5 to 13 MHz and the other 4 to 9 MHz. Comparison of spectral estimation methods showed that the Welch method provided spectra yielding more accurate and precise backscatter coefficient and scatterer size estimations than spectra computed by applying rectangular, Hanning, or Hamming windows and much reduced computational load than if using the multitaper method. For small echo signal data block sizes, moderate improvements in scatterer size estimations were obtained using a multitaper method, but this significantly increases the computational burden. It is critical to average power spectra from lateral A-lines for the improvement of scatterer size estimation. Averaging approximately 10 independent A-lines laterally with an axial window length 10 times the center frequency wavelength optimized trade-offs between spatial resolution and the variance of scatterer size estimates. Applying the concept of a time-bandwidth product, this suggests using analysis blocks that contain at least 30 independent samples of the echo signal. The estimation accuracy and precision depend on the ka range where k is the wave number and a is the effective scatterer size. This introduces a region-of-interest depth dependency to the accuracy and precision because of preferential attenuation of higher frequency sound waves in tissue-like media. With the 5 to 13 MHz, transducer ka ranged from 0.5 to 1.6 for scatterers in the test phantom, which is a favorable range, and the accuracy and precision of scatterer size estimations were both within ~5% using optimal analysis block dimensions. When the 4- to 9-MHz transducer was used, the ka value ranged from 0.3 to 0.8 to 1.1 for the experimental conditions, and the accuracy and precision were found to be ~10% and 10% to 25%, respectively. Although the experiments were done with 2 specific models of transducers on the test phantom, the results can be generalized to similar clinical arrays available from a variety of manufacturers and/or for different size of scatterers with similar ka range.

Tumor targeting studies using metallic nanoparticles (NPs) have shown that the enhanced permeability and retention effect may not be sufficient to deliver the amount of intratumoral and intracellular NPs needed for effective in vivo radiosensitization. This work describes a pH-Low Insertion Peptide (pHLIP) targeted theranostic agent to enable image-guided NP-enhanced radiotherapy using a clinically feasible amount of injected NPs. Conventional gadolinium (Gd) NPs were conjugated to pHLIPs and evaluated in vitro for radiosensitivity and in vivo for mouse MRI. Cultured A549 human lung cancer cells were incubated with 0.5 mM of pHLIP-GdNP or conventional GdNP. Mass spectrometry showed 78-fold more cellular Gd uptake with pHLIP-GdNPs, and clonogenic survival assays showed 44% more enhanced radiosensitivity by 5 Gy irradiation with pHLIP-GdNPs at pH 6.2. In contrast to conventional GdNPs, MR imaging of tumor-bearing mice showed pHLIP-GdNPs had a long retention time in the tumor (>9 h), suitable for radiotherapy, and penetrated into the poorly-vascularized tumor core. The Gd-enhanced tumor corresponded with low-pH areas also independently measured by an in vivo molecular MRI technique. pHLIPs actively target cell surface acidity from tumor cell metabolism and deliver GdNPs into cells in solid tumors. Intracellular delivery enhances the effect of short-range radiosensitizing photoelectrons and Auger electrons. Because acidity is a general hallmark of tumor cells, the delivery is more general than antibody targeting. Imaging the in vivo NP biodistribution and more acidic (often more aggressive) tumors has the potential for quantitative radiotherapy treatment planning and pre-selecting patients who will likely benefit more from NP radiation enhancement.