Python代码实现：基于scikit-image和NumPy计算PRO-AUC指标

Python代码实现：基于scikit-image和NumPy计算PRO-AUC指标

本文提供了一个Python代码示例，用于计算图像分割任务中的PRO-AUC（Per-Region Overlap - Area Under the Curve）指标。该指标可以有效评估分割模型在不同阈值下的性能。

代码pythonimport numpy as npfrom skimage import measurefrom sklearn.metrics import auc

def cal_pro_score(masks, amaps, max_step=200, expect_fpr=0.3): ''' 计算PRO-AUC分数。

参数:        masks: 真实掩码，形状为(N, H, W)，N为样本数量，H和W为图像高度和宽度。        amaps: 预测的激活映射，形状与masks相同。        max_step: 阈值范围内的步数。        expect_fpr: 期望的假阳性率。

返回值:        PRO-AUC分数。    '''    binary_amaps = np.zeros_like(amaps, dtype=bool)    min_th, max_th = amaps.min(), amaps.max()    delta = (max_th - min_th) / max_step    pros, fprs, ths = [], [], []        tp_pixels_sum = 0  # 记录总的tp_pixels总和    fp_pixels_sum = 0  # 记录总的fp_pixels总和        for th in np.arange(min_th, max_th, delta):        binary_amaps[amaps <= th], binary_amaps[amaps > th] = 0, 1                pro = []        tp_pixels_sum_th = 0  # 记录当前阈值下的tp_pixels总和        fp_pixels_sum_th = 0  # 记录当前阈值下的fp_pixels总和                for binary_amap, mask in zip(binary_amaps, masks):            for region in measure.regionprops(measure.label(mask)):                tp_pixels = binary_amap[region.coords[:, 0], region.coords[:, 1]].sum()                tp_pixels_sum += tp_pixels                tp_pixels_sum_th += tp_pixels  # 累加tp_pixels到总和                pro.append(tp_pixels / region.area)                    inverse_masks = 1 - masks            fp_pixels = np.logical_and(inverse_masks, binary_amaps).sum()            fp_pixels_sum += fp_pixels            fp_pixels_sum_th += fp_pixels  # 累加fp_pixels到总和                fpr = fp_pixels_sum_th / inverse_masks.sum()        pro_mean = np.array(pro).mean()                pros.append(pro_mean)        fprs.append(fpr)        ths.append(th)        pros, fprs, ths = np.array(pros), np.array(fprs), np.array(ths)    idxes = fprs < expect_fpr    fprs = fprs[idxes]    fprs = (fprs - fprs.min()) / (fprs.max() - fprs.min())    pro_auc = auc(fprs, pros[idxes])        print('fpr:', fprs)    print('TP:', tp_pixels_sum)    print('FP:', fp_pixels_sum)        return pro_a