Introduction

This is a simple framework for training neural networks to detect nodules in CT images. Training requires a json file (e.g. here) containing a list of CT images and the bounding boxes in each image. The model combines both CNN model and LSTM unit. The algorithm here is mainly refered to paper End-to-end people detection in crowded scenes.

The deep learning framewoek is based on TensorFlow(version 1.0.0) and some coding ideas are forked from the TensorBox project. Here I show heartfelt gratefulness. About nodule classfication method based on CNN transfer learning, you can refer to this paper.

Dicom and XML parsing

Parsing the lidc XML annotation and dicom files，see pylung and this blog.

>>> import dicom
>>> f = dicom.read_file('000001.dcm')
>>> print f
(0008, 0005) Specific Character Set              CS: 'ISO_IR 100'
(0008, 0008) Image Type                          CS: ['ORIGINAL', 'PRIMARY', 'AXIAL']
(0008, 0016) SOP Class UID                       UI: CT Image Storage
(0008, 0018) SOP Instance UID                    UI: 1.3.6.1.4.1.14519.5.2.1.6279.6001.143451261327128179989900675595
(0008, 0020) Study Date                          DA: '20000101'
(0008, 0021) Series Date                         DA: '20000101'
(0008, 0022) Acquisition Date                    DA: '20000101'
(0008, 0023) Content Date                        DA: '20000101'
(0008, 0024) Overlay Date                        DA: '20000101'
(0008, 0025) Curve Date                          DA: '20000101'
(0008, 002a) Acquisition DateTime                DT: '20000101'
(0008, 0030) Study Time                          TM: ''
(0008, 0032) Acquisition Time                    TM: ''
(0008, 0033) Content Time                        TM: ''
(0008, 0050) Accession Number                    SH: '2819497684894126'
(0008, 0060) Modality                            CS: 'CT'
(0008, 0070) Manufacturer                        LO: 'GE MEDICAL SYSTEMS'
(0008, 0090) Referring Physician Name            PN: ''
(0008, 1090) Manufacturer Model Name             LO: 'LightSpeed Plus'
(0008, 1155) Referenced SOP Instance UID         UI: 1.3.6.1.4.1.14519.5.2.1.6279.6001.675906998158803995297223798692
(0010, 0010) Patient Name                        PN: ''
(0010, 0020) Patient ID                          LO: 'LIDC-IDRI-0001'
(0010, 0030) Patient Birth Date                  DA: ''
(0010, 0040) Patient Sex                         CS: ''
(0010, 1010) Patient Age                         AS: ''
(0010, 21d0) Last Menstrual Date                 DA: '20000101'
(0012, 0062) Patient Identity Removed            CS: 'YES'
(0012, 0063) De-identification Method            LO: 'DCM:113100/113105/113107/113108/113109/113111'
(0013, 0010) Private Creator                     LO: 'CTP'
(0013, 1010) Private tag data                    LO: 'LIDC-IDRI'
(0013, 1013) Private tag data                    LO: '62796001'
(0018, 0010) Contrast/Bolus Agent                LO: 'IV'
(0018, 0015) Body Part Examined                  CS: 'CHEST'
(0018, 0022) Scan Options                        CS: 'HELICAL MODE'
(0018, 0050) Slice Thickness                     DS: '2.500000'
(0018, 0060) KVP                                 DS: '120'
(0018, 0090) Data Collection Diameter            DS: '500.000000'
(0018, 1020) Software Version(s)                 LO: 'LightSpeedApps2.4.2_H2.4M5'
(0018, 1100) Reconstruction Diameter             DS: '360.000000'
(0018, 1110) Distance Source to Detector         DS: '949.075012'
(0018, 1111) Distance Source to Patient          DS: '541.000000'
(0018, 1120) Gantry/Detector Tilt                DS: '0.000000'
(0018, 1130) Table Height                        DS: '144.399994'
(0018, 1140) Rotation Direction                  CS: 'CW'
(0018, 1150) Exposure Time                       IS: '570'
(0018, 1151) X-Ray Tube Current                  IS: '400'
(0018, 1152) Exposure                            IS: '4684'
(0018, 1160) Filter Type                         SH: 'BODY FILTER'
(0018, 1170) Generator Power                     IS: '48000'
(0018, 1190) Focal Spot(s)                       DS: '1.200000'
(0018, 1210) Convolution Kernel                  SH: 'STANDARD'
(0018, 5100) Patient Position                    CS: 'FFS'
(0020, 000d) Study Instance UID                  UI: 1.3.6.1.4.1.14519.5.2.1.6279.6001.298806137288633453246975630178
(0020, 000e) Series Instance UID                 UI: 1.3.6.1.4.1.14519.5.2.1.6279.6001.179049373636438705059720603192
(0020, 0010) Study ID                            SH: ''
(0020, 0011) Series Number                       IS: '3000566'
(0020, 0013) Instance Number                     IS: '80'
(0020, 0032) Image Position (Patient)            DS: ['-166.000000', '-171.699997', '-207.500000']
(0020, 0037) Image Orientation (Patient)         DS: ['1.000000', '0.000000', '0.000000', '0.000000', '1.000000', '0.000000']
(0020, 0052) Frame of Reference UID              UI: 1.3.6.1.4.1.14519.5.2.1.6279.6001.229925374658226729607867499499
(0020, 1040) Position Reference Indicator        LO: 'SN'
(0020, 1041) Slice Location                      DS: '-207.500000'
(0028, 0002) Samples per Pixel                   US: 1
(0028, 0004) Photometric Interpretation          CS: 'MONOCHROME2'
(0028, 0010) Rows                                US: 512
(0028, 0011) Columns                             US: 512
(0028, 0030) Pixel Spacing                       DS: ['0.703125', '0.703125']
(0028, 0100) Bits Allocated                      US: 16
(0028, 0101) Bits Stored                         US: 16
(0028, 0102) High Bit                            US: 15
(0028, 0103) Pixel Representation                US: 1
(0028, 0120) Pixel Padding Value                 US: 63536
(0028, 0303) Longitudinal Temporal Information M CS: 'MODIFIED'
(0028, 1050) Window Center                       DS: '-600'
(0028, 1051) Window Width                        DS: '1600'
(0028, 1052) Rescale Intercept                   DS: '-1024'
(0028, 1053) Rescale Slope                       DS: '1'
(0038, 0020) Admitting Date                      DA: '20000101'
(0040, 0002) Scheduled Procedure Step Start Date DA: '20000101'
(0040, 0004) Scheduled Procedure Step End Date   DA: '20000101'
(0040, 0244) Performed Procedure Step Start Date DA: '20000101'
(0040, 2016) Placer Order Number / Imaging Servi LO: ''
(0040, 2017) Filler Order Number / Imaging Servi LO: ''
(0040, a075) Verifying Observer Name             PN: 'Removed by CTP'
(0040, a123) Person Name                         PN: 'Removed by CTP'
(0040, a124) UID                                 UI: 1.3.6.1.4.1.14519.5.2.1.6279.6001.335419887712224178340067932923
(0070, 0084) Content Creator's Name              PN: ''
(0088, 0140) Storage Media File-set UID          UI: 1.3.6.1.4.1.14519.5.2.1.6279.6001.211790042620307056609660772296
(7fe0, 0010) Pixel Data                          OW: Array of 524288 bytes

Training

First, install TensorFlow from source or pip (NB: source installs currently break threading on 0.11). Then run the training script CNN_LSTM/run.sh.Note that running on your own dataset should require modifying the CNN_LSTM/hypes/\*.json file.

Evaluation

There are two options for evaluation, an ipython notebook and a python script.

• some test results

• precision-recall curve

Other

• Anatomy map

• Parenchyma extraction

• Pulmonary nodules form

Reference

• [Dataset] The Lung Image Database Consortium image collection(LIDC-IDRI)
• [Paper] End-to-end people detection in crowded scenes
• [Paper] Deep Convolutional Neural Networks for Computer-Aided Detection: CNN Architectures, Dataset Characteristics and Transfer Learning
• [Paper] An overview of classification algorithms for imbalanced datasets
• [Blog] LIDC-IDRI肺结节公开数据集Dicom和XML标注详解
• [Blog] 机器学习之分类性能度量指标 : ROC曲线、AUC值、正确率、召回率
• [简书] LIDC-IDRI肺结节Dicom数据集解析与总结
• [简书] LIDC-IDRI肺结节公开数据集Dicom和XML标注详解
• [简书] 医疗CT影像肺结节检测参考项目(附论文)
• [简书] 如何应用Python处理医学影像学中的DICOM信息
• [简书] CT图像肺结节识别算法调研 — CNN篇
• [简书] 吕乐：面向医学图像计算的深度学习与卷积神经网络（转）