data_prepare.py

import os
import random
import pandas as pd


def get_training_batch(root):
        """
        the proportional distribution of training data across each batch for classification,
        disease localization, report generation, and segmentation tasks to be 0.15/0.2/0.5/0.15
        """
        classification_label = ['Atelectasis', 'Calcification of the Aorta', 'Cardiomegaly', 'Consolidation', 'Edema', \
            'Emphysema', 'Enlarged Cardiomediastinum', 'Fibrosis', 'Fracture', 'Hernia', 'Infiltration', 'Lung Lesion', \
            'Lung Opacity', 'Mass', 'No Finding', 'Nodule', 'Pleural Effusion', 'Pleural Other', 'Pleural Thickening', \
            'Pneumomediastinum', 'Pneumonia', 'Pneumoperitoneum', 'Pneumothorax', 'Subcutaneous Emphysema', 'Support Devices', 'Tortuous Aorta']

        batch_size = 256
        cla_num = int(batch_size * 0.15)
        loc_num = int(batch_size * 0.2)
        report_num = int(batch_size * 0.5)
        seg_num = batch_size - (cla_num + loc_num + report_num)

        read = lambda x, y : pd.read_csv(x, sep='\t', header=None, chunksize=y)

        ### classification and report generation
        # Instruction: 
        #   what disease does this image have?
        #   is {} in this image?
        mimic = f'{root}/MIMIC_classification_report-generation_train.tsv'
        mimic_chunck = read(mimic, max(cla_num, report_num))
        cla_mimic_info = []
        report_mimic_info = []
        for chunck in mimic_chunck:
            for info in chunck.values.tolist():
                report = info[1]
                label = info[2]
                dicom_id = info[-1]
                if len(report_mimic_info) < report_num:
                    report_mimic_info.append(
                        ['describe the image', report, dicom_id, 'report generation']
                    )
                if len(cla_mimic_info) < int(cla_num*0.6):
                    if random.randint(0, 1):
                        cur_info = ['what disease does this image have?', f"there are {', '.join(label.split('&&'))}", dicom_id, 'classification']
                    else:
                        vqa_label = classification_label[random.randint(0, len(classification_label)-1)]
                        if vqa_label in label:
                            cur_info = [f'Is {vqa_label} in this image?', f'yes, there is {vqa_label}.', dicom_id, 'classification']
                        else:
                            cur_info = [f'Is {vqa_label} in this image?', f'no {vqa_label}.', dicom_id, 'classification']
                    cla_mimic_info.append(cur_info)
            break
        del mimic_chunck

        def organize_data(file, chunck_size, task, instruction, label_index, image_index, instruction_index=None, label_format=None):
            res = []
            chunck_size = max(chunck_size, 1)
            chuncks = read(file, chunck_size)
            for chunck in chuncks:
                for info in chunck.values.tolist():
                    if instruction_index is not None:
                        instruction = instruction.format(info[instruction_index])
                        ans = info[label_index]
                    elif label_format is not None:
                        label_ans_list = label_format(info[label_index])
                        label_ans = label_ans_list[random.randint(0, len(label_ans_list)-1)].split(',')
                        if len(label_ans) == 1:
                            label, ans = label_ans_list[0].split(',')[0], label_ans[0]
                        else:
                            label, ans = label_ans
                        label = label.strip()
                        instruction =  instruction.format(label)
                        ans = ans.strip()
                    if len(res) < chunck_size:
                        res.append(
                            [instruction, ans, info[image_index], task]
                        )
                break
            return res

        ### classification: severity
        # Instruction: 
        #   what is the level of {}?
        mimic_severity = f'{root}/MIMIC_classification-severity_train.tsv'
        cla_sev_mimic_info = organize_data(
            mimic_severity, int(cla_num*0.2), 'classification_sev', 'what is the level of {}?', 1, -1, label_format=lambda x:x.split('&&')
        )

        ### classification: location
        # Instruction: 
        #   where is {}?
        mimic_location = f'{root}/MIMIC_classification-location_train.tsv'
        cla_loc_mimic_info = organize_data(
            mimic_location, int(cla_num*0.2), 'classification_loc', 'where is {}?', 1, -1, label_format=lambda x:x.split('&')
        )

        ### localization
        # Instruction: 
        #   give the accurate bbox of {}.
        chestX_det = f'{root}/ChestX_Det_localization.tsv'
        chestX_det_info = organize_data(
            chestX_det, loc_num, 'localization', 'where is {}?', 2, -1, instruction_index=1, 
        )

        ### segmentation
        # Instruction: 
        #   describe the image.
        cheXmask_heart = f'{root}/CheXmask_heart_segmentation.tsv'
        cheXmask_heart_info = organize_data(
            cheXmask_heart, seg_num, 'segmentation', 'please segment the {} from the given image.', 2, -1, instruction_index=1
        )

        batch_info = cla_mimic_info + report_mimic_info + cla_loc_mimic_info + cla_sev_mimic_info + chestX_det_info + cheXmask_heart_info
        random.shuffle(batch_info)
        batch_df = pd.DataFrame(batch_info)
        return batch_df


if __name__ == '__main__':
    root = ''
    get_training_batch(root)