Update paper.md

Paper/paper.md

@@ -50,9 +50,9 @@ Towards the need for automated and precise AI-based analysis of medical images,
 
 # Statement of need
 
-The growing need for automated and robust analysis of medical images has driven the adoption of AI-based methods that often use DICOM images and RT structures as masks. However, the effectiveness of these AI approaches can vary due to differences in data sources and conversion techniques [1-3]. The DICOM standard includes the "radiotherapy structure set (RT-Struct)" object to facilitate the transfer of patient structures and related information, focusing on regions of interest and dose reference points.
+The growing need for automated and robust analysis of medical images has driven the adoption of AI-based methods that often use DICOM images and RT structures as masks. However, the effectiveness of these AI approaches can vary due to differences in data sources and conversion techniques [@Whybra2023-en][@Yousefirizi2023-ax][@Rufenacht2023-as]. The DICOM standard includes the "radiotherapy structure set (RT-Struct)" object to facilitate the transfer of patient structures and related information, focusing on regions of interest and dose reference points.
 
-Despite the availability of tools for converting DICOM images and RT-Structures into other formats [3,4], integrating auto-segmentation solutions using deep learning in clinical environments is rare due to the lack of open-source frameworks that handle DICOM RT-Structure sets effectively. Software packages like dcmrtstruct2nii, DicomRTTool [4], and PyRaDiSe [3] provide necessary functionalities, while frameworks like TorchIO [5] and MONAI [6] face limitations in processing DICOM RT-structure data. Research has shown that variations in mask generation methods affect patient clustering and radiomic-based modeling in multi-center studies [1,2].
+Despite the availability of tools for converting DICOM images and RT-Structures into other formats [@Rufenacht2023-as][@Anderson2021-fp], integrating auto-segmentation solutions using deep learning in clinical environments is rare due to the lack of open-source frameworks that handle DICOM RT-Structure sets effectively. Software packages like dcmrtstruct2nii, DicomRTTool [@Anderson2021-fp], and PyRaDiSe [@Rufenacht2023-as] provide necessary functionalities, while frameworks like TorchIO [@Perez-Garcia2021-jf] and MONAI [@Creators_The_MONAI_Consortium_undated-or] face limitations in processing DICOM RT-structure data. Research has shown that variations in mask generation methods affect patient clustering and radiomic-based modeling in multi-center studies [@Whybra2023-en][@Yousefirizi2023-ax].
 
 To address these challenges, we developed RT-utils, a specialized Python library designed to enhance the efficiency of manipulating RT-Structures. This tool aims to optimize workflows, simplify the handling of medical imaging data, and provide a comprehensive solution for researchers. RT-utils offers advanced techniques to convert expert-provided contours and AI tool output masks to RT-struct format, making them suitable for clinical workflows.
 
@@ -91,7 +91,7 @@ The add_roi method requires the ROI mask to follow a specific format: it should
 RT-utils spans a diverse range of technical capabilities such as Creating new RT Structs, Adding to existing RT Structs, loading an existing RT Struct contour as a mask and Merging two existing RT Structs. Rt-utils also has the parameter use_pin_hole is a Boolean value that is initially set to false. When enabled (set to true), it erases lines within a mask, allowing each distinct region in an image to be enclosed by a single contour instead of having nested contours. This feature is useful when working with RT-Struct viewers that do not support nested contours or contours with holes. These capabilities extend to various applications, offering accelerated development of deep learning models through standardized inputs. It facilitates the integration of RT-Struct data into computational analyses and image processing pipelines (e.g. radiomics and AI), contributing to the efficiency of medical image analysis. Moreover, the toolkit supports a smooth transition from predictive models to clinical workflows, enhancing the practical utility of automated segmentation. In essence, RT-utils not only simplifies the curation of RT-Struct data but also empowers users with versatile tools for interfacing with standard formats, thereby facilitating advanced medical image analysis and model integration. RT-utils is confined to a straightforward 2D-based conversion algorithm. This limitation might generate a synthetic appearance of RT contour i.e. pixelated contours which could potentially impede the acceptance of generated RT contours within clinical environments and in our future efforts this issue will be addressed.
 
 # Real-world Example
-For comparing the effects of different RT-Struct conversion methods, we investigated the RT-utils tool, dcmrtstruct2nii (https://github.com/Sikerdebaard/dcmrtstruct2nii) and the built-in tools from LIFEx [7] and 3D Slicer [8]. We implemented the conversion technique and conducted a comparison of the NIfTI ground truth files. The level of agreement observed between RT-utils and LIFEx surpasses that of other techniques. The mean absolute errors with respect to RT-utils are shown on sagittal and coronal masks. (Figures 1). The visual inspection of an example of converted masks overlaid on PET scans using different techniques are shown in Figures 1. 
+For comparing the effects of different RT-Struct conversion methods, we investigated the RT-utils tool, dcmrtstruct2nii (https://github.com/Sikerdebaard/dcmrtstruct2nii) and the built-in tools from LIFEx[@Nioche2018-ct] and 3D Slicer [@Fedorov2012-ax]. We implemented the conversion technique and conducted a comparison of the NIfTI ground truth files. The level of agreement observed between RT-utils and LIFEx surpasses that of other techniques. The mean absolute errors with respect to RT-utils are shown on sagittal and coronal masks. (Figures 1). The visual inspection of an example of converted masks overlaid on PET scans using different techniques are shown in Figures 1. 
 
 ![The visual inspection of an example of converted masks overlaid on PET scans using different techniques](https://github.com/user-attachments/assets/139bfba0-d25c-4373-8cf1-c603eeae1d6f)